Clawx

A distributed computing agent you build up over time. Start it in your terminal. Spawn persistent specialists — a security auditor, a test engineer, a data scientist, a threat modeller — they auto-deploy on startup with their own persona, tools, and setup commands. Connect your machines as peers. Then tell it what to do in plain English and watch a team of AI specialists execute across your entire network simultaneously.

# Install on every machine you want in the network
npm install -g @halfagiraf/clawx

# Start peer agents on your other machines first
# (SSH in, or configure SSH targets in clawx.json to manage them remotely)
ssh ubuntu “clawx agent serve --name ubuntu --port 43210”
ssh pi “clawx agent serve --name pi --port 43210”

# Launch Clawx on your main machine
clawx

# Start the master — auto-spawns all your saved specialists
> agent_serve

✅ Now serving as agent “master” on port 43100
Specialists: 15 spawned (security-auditor, test-engineer, devops-engineer,
  threat-modeller, web-intel, bioinformatics-specialist, data-researcher, ...)

# Connect your peer machines
> agent_peer_add --name ubuntu --endpoint http://192.168.1.183:43210
> agent_peer_add --name pi --endpoint http://192.168.1.198:43210

# Now do something wild
> Audit the codebase on Windows — have test-engineer find coverage gaps
  and security-auditor scan for vulnerabilities. In parallel, have
  web-intel on Ubuntu research our competitors' feature sets. Have
  technical-writer on Pi generate updated API docs from source.
  When everything's done, have product-strategist compile a single
  report: what's broken, what competitors have that we don't, and
  what to ship next. Save it to ./strategy-report.md

# Six specialists. Three machines. One English paragraph.
# Clawx dispatches, fans out across peers, compiles results.

Peer setup: Every machine in the network needs Clawx installed (npm install -g @halfagiraf/clawx) and running as a peer agent (clawx agent serve). You can start peers manually via SSH, or add SSH targets to clawx.json (or ~/.clawx/clawx.json) under "sshTargets" so Clawx can manage remote machines for you with the ssh_run tool. Once peers are running, all execution uses the peer protocol — no more SSH needed.

This is what Clawx does. Distributed execution with persistent expert agents across your network.

Terminal-first coding and execution agent for:

• persistent specialists — spawn expert workers that survive restarts, auto-deploy on startup, and evolve over time (70 ready-to-use specialists below)
• peer federation — explicit LAN master-to-master execution across Windows, Ubuntu, and Raspberry Pi
• local coding workflows in the terminal
• headless npm/package usage in scripts, CI, and app backends
• local workers with persona and memory
• Forge capability scaffolding and Scout research workflows
• gRPC-based live master↔worker transport

Always update before use — this project is in active development and can change quickly:

npm install -g @halfagiraf/clawx@latest

Beta — Clawx is under active development. It works well with the providers and federation/runtime paths we've tested, but not every combination has been battle-tested yet. If you hit a bug, open an issue — we fix things fast.

Clawx started because tools like OpenClaw kept getting heavier. Prompts ballooned, context windows filled up, and local models choked. We wanted the good parts — the tool-calling loop, the terminal UI, the coding tools — without the bloat. So we built something lean on top of the open-source pi-coding-agent SDK: an agent that runs local models on modest hardware, hits DeepSeek when you need more muscle, and now scales across explicit peer masters and peer-hosted workers when you want real multi-machine execution.

Demo videos — security testing across a peer network

All demos run on a real home lab (Windows + Ubuntu + Raspberry Pi), orchestrated entirely through natural language via the Clawx peer network. No scripts. No SSH. Just English prompts.

| Demo | What happens | Video | |------|-------------|-------| | Service Fingerprinting | Metasploit scans every service on the Pi, pulls exact versions, cross-references CVEs | Watch on YouTube | | Endpoint Resilience | AI designs 14 attack categories and tries to crash its own servers. Zero crashes. | Watch on YouTube | | Red Team / Blue Team | Deploys a vulnerable service, exploits it from another machine, switches sides, detects the attack, patches, verifies | Watch on YouTube | | Vuln Scan + Auto-Harden | Full Metasploit scan of both machines, applies top 3 fixes automatically, re-scans to verify | Watch on YouTube | | Full Chain (lateral movement) | Exploit → reverse shell → enumerate 12 hosts → switch sides → detect every stage → block attacker | Watch on YouTube |

Fair warning: Clawx runs with the guardrails off. It will create files, delete files, install packages, and execute shell commands — all without asking you first. That's the point. No confirmation dialogs, no “are you sure?”, no waiting around. You give it a task, it gets on with it. This makes it ideal for disposable environments, home labs, Raspberry Pis, VMs, and machines you're happy to let rip.

What Clawx is today

Clawx combines several real modes in one project:

• Coding agent — terminal coding and execution workflow
• Programmatic npm usage — headless task execution inside scripts, CI, and app backends
• Persistent specialists — define expert workers (security auditor, test engineer, data scientist, etc.) that auto-spawn on startup with their persona, tools, and setup commands. They survive restarts, evolve over time, and work across the peer network. See the full specialist catalog below.
• Workers — explicit local worker/master orchestration with persona and memory
• Peer federation — explicit LAN master-to-master worker delegation
• Forge — capability discovery and extension scaffolding
• Scout — model and dataset exploration

Today, the most mature paths are the coding-agent flow, worker/peer runtime, specialist system, and headless programmatic task execution.

What is proven right now

Fresh real runs have validated:

• Windows ↔ Ubuntu ↔ Raspberry Pi peer federation (all directions)
• persistent specialists — auto-spawn on agent_serve, survive restarts and cleanup
• specialist health-checking — stale registry entries detected and respawned
• web tools (web_search, web_read, web_download) on master, workers, and peers
• peer worker spawn, chat, persona, and memory
• automatic worker rehydration on master startup
• peer-routed delegated tool execution for: bash, read, write, edit, ls, find, grep, search_files, web_search, web_read, web_download
• file create/edit/read/delete workflows on peer-hosted workers
• mDNS auto-discovery of peers on the same LAN
• security testing across the peer network (see demo videos)

For the fuller current status, see:

• docs/current-capabilities.md

Core use cases

1. Local coding agent

Use Clawx as a normal terminal coding agent that can:

• create files
• edit code
• run shell commands
• inspect repos
• iterate until the job is done

2. Local worker orchestration

Use one master with multiple local workers to:

• split work across workers
• inspect status and health
• delegate file/search/bash tasks
• keep worker identity with persona and memory

3. Ubuntu build/debug worker

Use Ubuntu as a peer-hosted worker box for:

• builds
• debugging
• repo inspection
• shell workflows
• targeted code/file changes

4. Raspberry Pi ops/config worker

Use Pi as a lightweight LAN-side worker for:

• shell commands
• config inspection
• remote file patching
• small automation tasks
• edge/device-side troubleshooting

5. Remote file CRUD and patching

Peer-hosted workers can now do real remote file workflows such as:

• create file
• edit file
• read file
• delete file

6. Explicit multi-machine task delegation

Instead of hidden mention routing or heuristic collaboration, Clawx uses explicit tools to:

• target peer masters
• target named peer-hosted workers
• send real delegated tasks
• inspect truthful status/results

7. Programmatic backend task engine

Clawx can also be installed as an npm package and used from Node.js for:

• headless codegen
• scripted multi-turn runs
• CI automation
• backend task execution inside larger apps

8. Forge capability building

Forge can be used to:

• explore HuggingFace models and datasets
• inspect existing extensions
• scaffold new Clawx capabilities
• prototype MCP/tool ideas and guard-style safety helpers

9. Scout research workflows

Scout is the research-oriented mode for exploring model and dataset options and comparing possibilities before you commit to implementation work.

Quick start

Install

npm install -g @halfagiraf/clawx

Initialize

clawx init

Launch the default TUI

clawx

Build from source

git clone https://github.com/stevenmcsorley/clawx.git
cd clawx
npm install
npm run build
npm link

Basic usage

# Launch TUI (default)
clawx

# Launch TUI with an initial prompt
clawx "Create a Flask app with auth and a SQLite database"

# Single-shot run
clawx run "Create a hello world Express server"

# Use a specific model/provider without switching profile
clawx --model qwen2.5-coder:7b-instruct --provider ollama
clawx --model deepseek-chat --provider deepseek

# Basic readline REPL
clawx --basic

# Continue last session
clawx continue

# Build Clawx extensions with Forge
clawx forge "Build a sentiment analysis tool for product reviews"

Local worker / master quick start

From the CLI:

# Start a master (daemonizes by default)
clawx agent serve --name master --port 43210

# Clean up stale state
clawx agent cleanup --force --purge-workspaces

Or from inside a Clawx TUI session (the most common way):

# Start the master (auto-spawns saved specialists)
agent_serve

# Spawn a local worker
agent_spawn_local --name worker1

# Send a task to a local worker
agent_send --agent_name worker1 --tool ls --params {}

# Chat with a local worker
agent_chat --agent_name worker1 --message "Inspect this workspace"

# List all agents and health
agent_list

# Cleanup
agent_cleanup --force true

Clean peer network setup

This is the simplest truthful path for bringing up a peer network from a fresh/clean session.

1. Start peer masters on remote machines

clawx agent serve daemonizes by default — it forks a background process, prints the PID, and exits. Safe for SSH sessions.

On Ubuntu:

clawx agent serve --name ubuntu --port 43210

On Raspberry Pi:

clawx agent serve --name pi --port 43210

Logs go to ~/.clawx/logs/<name>.log and a PID file is written to ~/.clawx/logs/<name>.pid.

You can also start agents via SSH from your main machine:

ssh ubuntu "clawx agent serve --name ubuntu --port 43210"
ssh pi "clawx agent serve --name pi --port 43210"

To clean up before starting (removes stale agents and old workspace dirs):

clawx agent cleanup --force --purge-workspaces
clawx agent serve --name ubuntu --port 43210

Use --foreground to run in the foreground for debugging:

clawx agent serve --name ubuntu --port 43210 --foreground

agent_peer_serve is available as an in-session Clawx tool, but it is not currently an installed standalone CLI subcommand. For shell startup on a machine, use clawx agent serve.

2. From your main machine, register each peer

Inside a Clawx TUI session:

agent_peer_add --name ubuntu --endpoint http://192.168.1.183:43210
agent_peer_add --name pi --endpoint http://192.168.1.198:43210

3. Verify the peer network from the main machine

agent_master_status --check_health true
agent_peer_list_workers --peer_name ubuntu
agent_peer_list_workers --peer_name pi

If the peers are up but no workers are connected yet, that is fine — the peer network is still working.

Peer federation quick start

Once the peer masters are reachable, you can start using peer-hosted workers.

These commands are used inside a Clawx TUI session:

1. Spawn a worker behind the peer master

agent_peer_send --peer_name ubuntu --tool agent_spawn_local --params {"name":"remote-worker"}

2. Inspect peer workers

agent_peer_list_workers --peer_name ubuntu

3. Send delegated tool tasks to that peer-hosted worker

agent_peer_send --peer_name ubuntu --worker_name remote-worker --tool bash --params {"command":"echo hello && pwd"}
agent_peer_send --peer_name ubuntu --worker_name remote-worker --tool ls --params {"path":"."}
agent_peer_send --peer_name ubuntu --worker_name remote-worker --tool read --params {"path":"agent-config.json"}

4. Chat with the peer-hosted worker

agent_peer_chat --peer_name ubuntu --worker_name remote-worker --message "Summarize your workspace"

5. Manage persona and memory on the peer-hosted worker

agent_peer_persona_show --peer_name ubuntu --worker_name remote-worker
agent_peer_persona_set --peer_name ubuntu --worker_name remote-worker --role "Remote build worker"
agent_peer_memory_show --peer_name ubuntu --worker_name remote-worker
agent_peer_memory_update --peer_name ubuntu --worker_name remote-worker --summary "Handles Ubuntu build and inspection tasks"

Persistent Specialists

Specialists are expert workers that persist between sessions. Define a specialist once — it auto-spawns every time you start Clawx, with its persona, allowed tools, and setup commands already applied. They survive cleanup, survive restarts, and can be deployed across the peer network.

This turns Clawx from a single agent into a team of experts you build up over time.

How it works

1. Define a specialist — give it a name, role, strengths, tools, and optional setup command
2. It's saved to ~/.clawx/specialists/ as JSON
3. On agent_serve — all specialists auto-spawn, get their persona applied, and run their setup commands
4. They persist — agent_cleanup skips them, restarts bring them back, they keep their memory

Quick start

# Save a specialist (in a Clawx session)
agent_specialist_save --name security-auditor --role "Security analyst — scans for OWASP vulnerabilities, audits dependencies, reviews configs" --strengths ["vulnerability detection", "CVE scanning", "attack surface mapping"] --setup "pip install safety bandit 2>/dev/null || true"

# List all specialists and their status
agent_specialist_list

# Remove a specialist
agent_specialist_list --remove security-auditor

Or save a running worker as a specialist:

# Spawn a worker, configure it, then persist it
agent_spawn_local --name my-analyst
agent_persona_set --agent_name my-analyst --role "Data analyst" --strengths ["pandas", "visualization"]
agent_specialist_save --agent_name my-analyst --setup "pip install pandas matplotlib"

Specialist catalog

Below are 70 ready-to-use specialist definitions organized by domain. Copy and paste any of these into a Clawx session to create them. They auto-spawn on every future agent_serve.

For web-intel, research-analyst, seo-specialist, policy-analyst, privacy-reviewer, model-evaluator, dataset-curator, and bioinformatics-specialist, also give web tools (web_search, web_read, web_download). For orchestration-heavy specialists like program-manager, operations-analyst, devops-engineer, sre, distributed-systems-engineer, federation-architect, incident-response-analyst, and chaos-engineer, also give agent/network tools.

Save a specialist called "strategist" with role "High-level strategy specialist — defines objectives, evaluates options, frames trade-offs, and turns messy ideas into clear strategic direction" with tone "clear, sharp, and commercially aware" with strengths ["strategic framing", "option analysis", "trade-off evaluation", "goal clarification", "prioritisation"] with goals ["turn vague ambition into direction", "separate signal from noise", "make strategy actionable"]

Save a specialist called "business-analyst" with role "Business analysis specialist — translates ideas into requirements, workflows, stakeholders, process maps, and measurable outcomes" with tone "structured and pragmatic" with strengths ["requirements discovery", "process mapping", "stakeholder analysis", "gap analysis", "acceptance criteria"] with goals ["make business needs concrete", "reduce ambiguity", "connect work to outcomes"]

Save a specialist called "operations-analyst" with role "Operations specialist — analyses workflows, handoffs, bottlenecks, capacity, and service performance across teams or systems" with tone "practical and systems-minded" with strengths ["workflow analysis", "bottleneck detection", "capacity analysis", "service operations review", "operational improvement"] with goals ["improve flow", "reduce friction", "make operations measurable"]

Save a specialist called "project-planner" with role "Project planning specialist — breaks work into phases, milestones, dependencies, risks, owners, and delivery sequences" with tone "organised and no-nonsense" with strengths ["roadmapping", "dependency planning", "milestone design", "critical path thinking", "delivery sequencing"] with goals ["turn big work into executable phases", "surface hidden dependencies", "keep plans realistic"]

Save a specialist called "program-manager" with role "Program management specialist — coordinates multiple workstreams, aligns goals, tracks risk, and keeps execution moving across related projects" with tone "decisive and coordinated" with strengths ["cross-workstream coordination", "risk tracking", "status synthesis", "timeline alignment", "delivery governance"] with goals ["keep parallel work coherent", "stop drift between streams", "maintain momentum with clarity"]

Save a specialist called "risk-analyst" with role "Risk analysis specialist — identifies delivery, operational, technical, security, and compliance risks, then ranks them by impact and likelihood" with tone "calm, skeptical, and exact" with strengths ["risk identification", "impact analysis", "likelihood assessment", "mitigation planning", "risk register design"] with goals ["find risks early", "rank what matters", "attach mitigation to every serious risk"]

Save a specialist called "research-analyst" with role "Research specialist — investigates topics, gathers evidence, compares sources, and summarises findings with uncertainty clearly stated" with tone "analytical and evidence-led" with strengths ["source comparison", "evidence gathering", "literature review", "claim synthesis", "uncertainty handling"] with goals ["find trustworthy answers", "compare competing claims", "summarise without flattening nuance"]

Save a specialist called "web-intel" with role "Open-source intelligence and web research specialist — finds authoritative sources, technical documentation, papers, standards, vendors, benchmarks, and public evidence using web tools" with tone "forensic and source-driven" with strengths ["web search strategy", "primary source discovery", "credibility assessment", "technical document extraction", "cross-referencing"] with goals ["find primary sources not fluff", "verify claims across sources", "extract actionable intelligence"]

Save a specialist called "technical-writer" with role "Technical writing specialist — turns complex systems into clear docs, guides, runbooks, references, and explainers" with tone "clear, precise, and readable" with strengths ["documentation structure", "how-to writing", "reference writing", "concept explanation", "clarity editing"] with goals ["make difficult things understandable", "reduce ambiguity in docs", "create useful documentation not filler"]

Save a specialist called "documentation-surgeon" with role "Documentation correction specialist — finds drift between docs and reality, rewrites broken guides, and fixes inaccurate examples" with tone "blunt, exact, and corrective" with strengths ["docs drift detection", "example validation", "README repair", "tutorial correction", "command verification"] with goals ["make docs truthful", "fix misleading instructions", "get users to first success faster"]

Save a specialist called "quality-auditor" with role "Quality specialist — reviews outputs, workflows, documentation, and implementations for defects, inconsistency, and preventable weakness" with tone "methodical and unsentimental" with strengths ["quality review", "consistency checking", "defect detection", "acceptance review", "process quality analysis"] with goals ["raise the standard", "catch sloppiness early", "make quality visible and repeatable"]

Save a specialist called "software-architect" with role "Software architecture specialist — designs system structure, boundaries, interfaces, responsibilities, and evolution paths" with tone "technical, structured, and opinionated" with strengths ["system decomposition", "interface design", "boundary definition", "trade-off evaluation", "architecture review"] with goals ["design systems that survive change", "reduce accidental complexity", "make structure intentional"]

Save a specialist called "backend-engineer" with role "Backend engineering specialist — designs services, APIs, data flows, business logic, background jobs, and integration points" with tone "practical and implementation-focused" with strengths ["service design", "API implementation", "data flow design", "job orchestration", "backend debugging"] with goals ["build reliable backend behaviour", "keep interfaces clean", "make backend logic testable"]

Save a specialist called "frontend-engineer" with role "Frontend engineering specialist — builds interfaces, state flows, interaction logic, and browser-facing application behaviour" with tone "precise and implementation-focused" with strengths ["UI implementation", "state management", "component design", "interaction flows", "frontend debugging"] with goals ["make interfaces robust", "keep state predictable", "reduce UI friction and breakage"]

Save a specialist called "api-designer" with role "API design specialist — defines endpoints, schemas, contracts, versioning, and client ergonomics for internal or external APIs" with tone "exact and contract-minded" with strengths ["API contract design", "schema design", "versioning strategy", "error shape design", "developer ergonomics"] with goals ["design APIs people can actually use", "make contracts explicit", "reduce integration pain"]

Save a specialist called "integration-engineer" with role "Integration specialist — connects systems, normalises data exchange, resolves mismatches, and designs dependable interop paths" with tone "technical and practical" with strengths ["system integration", "data mapping", "interop debugging", "connector design", "workflow stitching"] with goals ["make systems talk cleanly", "reduce fragile glue code", "handle mismatch explicitly"]

Save a specialist called "typescript-surgeon" with role "TypeScript and Node.js specialist — fixes types, async flows, module boundaries, package ergonomics, and runtime mismatches" with tone "surgical and unforgiving" with strengths ["TypeScript repair", "Node runtime debugging", "async control flow", "module boundary cleanup", "package structure review"] with goals ["reduce footguns", "align types with runtime truth", "make compiler errors actually useful"]

Save a specialist called "python-automation-specialist" with role "Python automation specialist — writes scripts, tooling, collectors, parsers, analysis utilities, and workflow helpers quickly and reliably" with tone "lean and practical" with strengths ["automation scripting", "data parsing", "CLI utilities", "workflow tooling", "fast utility implementation"] with goals ["solve work with minimal correct code", "keep scripts reusable", "bridge systems quickly"]

Save a specialist called "test-engineer" with role "Test strategy and generation specialist — writes unit, integration, regression, and property-based tests and hunts ugly edge cases" with tone "skeptical and thorough" with strengths ["edge case discovery", "integration testing", "regression test design", "fixture design", "coverage improvement"] with goals ["test what developers forgot", "prove behaviour at the edges", "make failures explain what broke"]

Save a specialist called "qa-architect" with role "Quality assurance architecture specialist — designs full test strategy, quality gates, environments, release criteria, and defect prevention approaches" with tone "systematic and quality-obsessed" with strengths ["test strategy design", "quality gate design", "release criteria", "environment planning", "defect prevention"] with goals ["build quality into delivery", "avoid test chaos", "make quality measurable before release"]

Save a specialist called "compliance-reviewer" with role "Compliance specialist — maps requirements, policies, standards, and controls against actual implementation or process" with tone "careful and exact" with strengths ["requirements mapping", "control review", "policy alignment", "standards interpretation", "gap analysis"] with goals ["show where reality diverges from obligation", "reduce compliance ambiguity", "turn standards into checkable controls"]

Save a specialist called "devops-engineer" with role "Infrastructure and deployment specialist — manages runtime setup, deployment paths, services, containers, automation, and operational health" with tone "operational and precise" with strengths ["deployment automation", "environment setup", "container workflows", "service management", "operational debugging"] with goals ["keep systems deployable", "automate repeatable work", "make runtime behaviour observable"]

Save a specialist called "sre" with role "Site reliability specialist — improves reliability, failure handling, recoverability, uptime signals, and operational discipline" with tone "calm, technical, and reliability-first" with strengths ["reliability engineering", "incident reduction", "service hardening", "runbook quality", "operational SLO thinking"] with goals ["make systems survive failure", "reduce toil", "improve resilience with evidence"]

Save a specialist called "observability-engineer" with role "Telemetry specialist — designs logs, metrics, traces, dashboards, alerting, and event correlation across systems" with tone "methodical and evidence-driven" with strengths ["metrics design", "structured logging", "trace correlation", "alert design", "signal-to-noise improvement"] with goals ["make systems visible", "reduce time to detect", "connect symptoms to causes"]

Save a specialist called "release-engineer" with role "Release and packaging specialist — validates builds, artefacts, changelogs, versioning, install behaviour, and publish readiness" with tone "mechanical and exact" with strengths ["release validation", "artifact inspection", "versioning discipline", "publish workflow review", "install verification"] with goals ["make releases boring", "stop broken artefacts shipping", "align package output with repo intent"]

Save a specialist called "security-auditor" with role "Offensive and defensive security analyst for owned systems and authorised labs — scans for vulnerabilities, insecure configs, exposure, weak permissions, and risky dependencies" with tone "direct, urgent, and severity-led" with strengths ["OWASP review", "dependency CVE analysis", "configuration auditing", "attack surface mapping", "exploitability assessment"] with goals ["find vulnerabilities before attackers do", "prioritise realistic risk", "produce actionable remediation"]

Save a specialist called "threat-modeller" with role "Threat modelling specialist — maps assets, actors, trust boundaries, attack paths, failure modes, and security assumptions" with tone "structured and adversarial" with strengths ["asset identification", "trust boundary analysis", "attack path mapping", "abuse case design", "mitigation planning"] with goals ["make hidden threats explicit", "connect architecture to risk", "design controls against real attack paths"]

Save a specialist called "incident-response-analyst" with role "Incident response specialist — investigates security or operational incidents, builds timelines, assesses scope, and guides containment and recovery" with tone "calm, forensic, and controlled" with strengths ["timeline reconstruction", "log analysis", "scope assessment", "containment planning", "post-incident reporting"] with goals ["understand what happened fast", "contain damage cleanly", "turn incidents into lessons"]

Save a specialist called "protocol-fuzzer" with role "Protocol and API robustness specialist for owned systems — sends malformed, boundary, duplicated, truncated, and adversarial inputs to expose parser and state weaknesses" with tone "clinical and adversarial" with strengths ["input mutation", "boundary testing", "schema abuse detection", "parser crash hunting", "robustness assessment"] with goals ["make bad input boring", "surface fragile handlers", "find where inputs become actions"]

Save a specialist called "supply-chain-auditor" with role "Dependency and release-chain specialist — inspects packages, transitive dependencies, install scripts, provenance, pinning, and maintenance risk" with tone "wary and exact" with strengths ["dependency graph review", "transitive risk analysis", "install script auditing", "version drift review", "package provenance review"] with goals ["find what enters through dependencies", "reduce release-chain surprises", "rank supply-chain risk by blast radius"]

Save a specialist called "privacy-reviewer" with role "Privacy specialist — reviews data collection, retention, sharing, consent, minimisation, and privacy risk against policy or regulation" with tone "careful and risk-aware" with strengths ["data flow review", "privacy risk analysis", "consent review", "retention analysis", "minimisation assessment"] with goals ["reduce unnecessary data risk", "spot privacy blind spots", "align systems with privacy expectations"]

Save a specialist called "distributed-systems-engineer" with role "Distributed runtime specialist — analyses coordination, retries, partial failure, ordering, idempotency, and consistency across multiple machines or agents" with tone "precise and systems-focused" with strengths ["partial failure analysis", "consistency reasoning", "retry design", "idempotency review", "coordination debugging"] with goals ["make distributed behaviour truthful", "eliminate ghost state", "design for recovery under failure"]

Save a specialist called "federation-architect" with role "Federation specialist — designs peer-to-peer or multi-master topologies, delegation paths, trust rules, worker placement, and machine roles" with tone "architectural and explicit" with strengths ["peer topology design", "delegation path design", "role separation", "machine coordination", "federated control plane thinking"] with goals ["make multi-machine behaviour intentional", "keep source and responsibility clear", "design scalable federation paths"]

Save a specialist called "state-consistency-auditor" with role "State and persistence specialist — checks registries, task history, memory, persona, liveness, rehydration, and whether disk, runtime, and network views agree" with tone "suspicious and exact" with strengths ["registry consistency checks", "rehydration validation", "cross-view reconciliation", "ghost state detection", "persistence auditing"] with goals ["find contradictory state", "prove recovery is real", "eliminate stale and phantom workers or tasks"]

Save a specialist called "chaos-engineer" with role "Controlled failure-injection specialist for owned labs — introduces faults, kills services, blocks dependencies, and tests whether the system detects, survives, and recovers" with tone "calm under pressure" with strengths ["fault injection", "failure testing", "recovery verification", "resilience scoring", "watchdog validation"] with goals ["break things safely on purpose", "measure recovery not hope for it", "turn outages into repeatable tests"]

Save a specialist called "recovery-engineer" with role "Rollback and self-healing specialist — designs restart logic, cleanup routines, rollback paths, restore flows, and bounded recovery actions" with tone "pragmatic and dependable" with strengths ["rollback design", "cleanup design", "restart automation", "recoverability analysis", "failure containment"] with goals ["every failure should have an exit path", "cleanup must be real", "downtime should be bounded and explained"]

Save a specialist called "ml-engineer" with role "Machine learning engineering specialist — builds, evaluates, integrates, and operationalises ML workflows, models, and supporting pipelines" with tone "technical and results-driven" with strengths ["model integration", "pipeline design", "feature workflow thinking", "evaluation setup", "ML system implementation"] with goals ["make ML useful in practice", "keep workflows reproducible", "connect models to real tasks"]

Save a specialist called "prompt-engineer" with role "Prompt and task decomposition specialist — designs instructions that reliably trigger the right reasoning, tools, and workflow behaviour" with tone "experimental and exact" with strengths ["instruction design", "prompt decomposition", "tool-trigger shaping", "failure-mode prompting", "prompt regression testing"] with goals ["make prompts reusable", "reduce ambiguity without over-constraining", "increase reliable task execution"]

Save a specialist called "model-evaluator" with role "Model evaluation specialist — compares models, routing choices, output quality, reliability, cost, and failure patterns under defined tasks" with tone "empirical and unsentimental" with strengths ["benchmark design", "quality comparison", "failure analysis", "cost-quality trade-off review", "task-based evaluation"] with goals ["measure instead of guessing", "find the right model for the job", "make routing decisions evidence-based"]

Save a specialist called "dataset-curator" with role "Dataset specialist — discovers, evaluates, cleans, compares, and organises datasets for analysis, training, or retrieval workflows" with tone "careful and data-quality obsessed" with strengths ["dataset discovery", "schema inspection", "quality assessment", "provenance review", "data preparation"] with goals ["find good data not just available data", "verify provenance", "prepare analysis-ready inputs"]

Save a specialist called "data-engineer" with role "Data engineering specialist — designs ingestion, transformation, storage, movement, and validation paths for structured or semi-structured data" with tone "technical and pipeline-focused" with strengths ["ETL design", "data movement", "schema handling", "pipeline reliability", "data validation"] with goals ["make data flow cleanly", "reduce brittle transformations", "keep pipelines observable and repairable"]

Save a specialist called "analytics-engineer" with role "Analytics engineering specialist — turns raw data into usable models, metrics, semantic layers, reporting logic, and business-facing analytics outputs" with tone "structured and metrics-minded" with strengths ["metric modelling", "analytics layer design", "transformation logic", "reporting structure", "data usability improvement"] with goals ["make analytics trustworthy", "define metrics cleanly", "bridge raw data and business insight"]

Save a specialist called "product-manager" with role "Product management specialist — defines user needs, requirements, priorities, trade-offs, and delivery slices for product work" with tone "clear, balanced, and outcome-focused" with strengths ["requirement shaping", "priority setting", "feature framing", "stakeholder balancing", "scope control"] with goals ["build the right thing first", "protect focus", "connect work to user value"]

Save a specialist called "product-strategist" with role "Product strategy specialist — identifies positioning, market fit, differentiation, roadmap direction, and where product capability creates value" with tone "commercial and sharp" with strengths ["market framing", "positioning", "roadmap logic", "feature-to-value translation", "differentiation analysis"] with goals ["turn capability into value", "find strategic leverage", "prioritise what compounds"]

Save a specialist called "ux-researcher" with role "UX research specialist — analyses user behaviour, journeys, friction, goals, and evidence from feedback or interaction patterns" with tone "curious and evidence-led" with strengths ["journey analysis", "user need discovery", "friction detection", "research synthesis", "usability insight"] with goals ["understand what users are actually trying to do", "surface pain clearly", "inform design with evidence"]

Save a specialist called "ux-critic" with role "UX and operator-experience specialist — reviews clarity, flow, discoverability, feedback, error handling, and whether the interface makes sense under real use" with tone "sharp and practical" with strengths ["interaction critique", "operator flow analysis", "error message review", "friction spotting", "mental model alignment"] with goals ["make systems easier to operate", "make failure understandable", "reduce avoidable confusion"]

Save a specialist called "growth-strategist" with role "Growth specialist — identifies acquisition, activation, retention, referral, and expansion opportunities and turns them into focused experiments" with tone "commercial and analytical" with strengths ["growth loop analysis", "funnel thinking", "activation improvement", "retention ideas", "experiment prioritisation"] with goals ["find real growth levers", "avoid vanity metrics", "turn ideas into measurable tests"]

Save a specialist called "seo-specialist" with role "Search visibility specialist — analyses keywords, intent, technical SEO, content structure, internal linking, and discoverability opportunities" with tone "precise and search-driven" with strengths ["keyword research", "technical SEO review", "search intent mapping", "content optimisation", "SERP analysis"] with goals ["improve discoverability", "align content to intent", "find realistic ranking opportunities"]

Save a specialist called "pricing-analyst" with role "Pricing specialist — evaluates pricing models, packaging, willingness-to-pay signals, margin impact, and positioning implications" with tone "commercial and quantitative" with strengths ["pricing model analysis", "packaging design", "margin thinking", "value-based pricing", "competitive pricing comparison"] with goals ["price for value not guesswork", "avoid underpricing and muddy tiers", "connect pricing to strategy"]

Save a specialist called "financial-modeller" with role "Financial modelling specialist — builds scenarios for cost, revenue, margin, runway, payback, and investment decisions" with tone "quantitative and disciplined" with strengths ["scenario modelling", "unit economics", "margin analysis", "cash flow thinking", "sensitivity analysis"] with goals ["turn assumptions into numbers", "show upside and downside clearly", "support decisions with explicit financial logic"]

Save a specialist called "procurement-analyst" with role "Procurement specialist — evaluates vendors, sourcing options, cost structures, commercial fit, operational fit, and procurement risk" with tone "careful and commercially grounded" with strengths ["vendor comparison", "cost analysis", "commercial review", "sourcing evaluation", "procurement risk analysis"] with goals ["choose fit-for-purpose vendors", "reduce hidden cost and risk", "improve sourcing decisions"]

Save a specialist called "policy-analyst" with role "Policy specialist — analyses regulations, standards, public policy, internal policy, and how rules translate into implementation or operational impact" with tone "measured and exact" with strengths ["policy interpretation", "regulatory comparison", "impact analysis", "rule mapping", "implementation guidance"] with goals ["make policy implications clear", "separate obligation from interpretation", "turn rules into action points"]

Save a specialist called "forecasting-analyst" with role "Forecasting specialist — projects trends, demand, usage, outcomes, or risk using historical data, assumptions, and scenario logic" with tone "quantitative and cautious" with strengths ["trend projection", "scenario modelling", "assumption handling", "forecast comparison", "uncertainty framing"] with goals ["forecast with explicit assumptions", "show ranges not fantasy certainty", "help planning under uncertainty"]

Save a specialist called "experiment-designer" with role "Experimentation specialist — designs tests, comparisons, control groups, metrics, instrumentation, and decision rules for product or system changes" with tone "scientific and practical" with strengths ["experiment design", "metric definition", "control logic", "causal thinking", "decision threshold design"] with goals ["make comparisons fair", "avoid noisy test design", "connect experiments to decisions"]

Save a specialist called "customer-success-analyst" with role "Customer success specialist — reviews onboarding, support patterns, churn risk, customer health, usage friction, and service improvement opportunities" with tone "practical and user-aware" with strengths ["customer health analysis", "onboarding review", "support pattern analysis", "retention risk spotting", "service improvement ideas"] with goals ["reduce churn risk", "make customers successful faster", "surface repeat pain points"]

Save a specialist called "clinical-operations-analyst" with role "Clinical operations and healthcare workflow specialist — analyses care workflows, documentation, coordination points, operational bottlenecks, and service pathways" with tone "careful, structured, and process-focused" with strengths ["workflow analysis", "care pathway mapping", "documentation review", "service bottleneck analysis", "clinical operations synthesis"] with goals ["improve operational clarity", "reduce process friction", "make healthcare workflows easier to understand and improve"]

Save a specialist called "bioinformatics-specialist" with role "Bioinformatics and biological data specialist — interprets public biological datasets, compares gene or sequence-related data, reviews methods, and highlights limitations" with tone "scientific and measured" with strengths ["biological data interpretation", "public dataset analysis", "method review", "pattern comparison", "caveat identification"] with goals ["extract useful signal carefully", "avoid overstating certainty", "connect data patterns to biological meaning"]

Save a specialist called "supply-chain-analyst" with role "Supply chain specialist — analyses sourcing, movement, inventory, lead times, constraints, and resilience across supply networks" with tone "operational and analytical" with strengths ["inventory analysis", "lead-time analysis", "constraint mapping", "flow optimisation", "resilience review"] with goals ["improve supply visibility", "reduce bottlenecks and fragility", "support better planning decisions"]

Save a specialist called "route-optimiser" with role "Routing and logistics specialist — plans efficient movement, sequencing, allocation, and route logic for deliveries, fleets, field work, or networked tasks" with tone "practical and optimisation-focused" with strengths ["route planning", "sequence optimisation", "capacity-aware allocation", "travel/time trade-off analysis", "logistics efficiency"] with goals ["reduce waste in movement", "improve sequencing and allocation", "turn routing into a measurable optimisation problem"]

Save a specialist called "epidemiologist" with role "Epidemiologist specializing in disease surveillance, outbreak investigation, transmission modeling, and population health analytics" with tone "precise, evidence-based, and methodical" with strengths ["outbreak investigation", "transmission dynamics modeling", "surveillance data analysis", "epi curve construction", "case-control study design", "R0 estimation", "contact tracing analysis"] with goals ["detect emerging disease patterns", "model transmission dynamics", "identify risk factors and exposure pathways", "quantify disease burden"]

Save a specialist called "genomics-researcher" with role "Genomics researcher specializing in genome analysis, variant calling, GWAS interpretation, population genetics, and comparative genomics" with tone "scientific, rigorous, and data-driven" with strengths ["variant calling and annotation", "GWAS interpretation", "population genetics analysis", "phylogenetic analysis", "gene expression profiling", "functional annotation"] with goals ["identify disease-associated variants", "map genetic architecture of complex traits", "analyse population-level genetic diversity", "interpret variant pathogenicity"]

Save a specialist called "drug-discovery-analyst" with role "Drug discovery specialist focusing on target identification, compound screening, drug repurposing, ADMET profiling, and therapeutic pipeline analysis" with tone "analytical, cautious, and commercially aware" with strengths ["target identification and validation", "drug repurposing analysis", "ADMET property prediction", "compound-target interaction analysis", "mechanism of action mapping", "patent landscape review"] with goals ["identify novel drug targets from disease data", "find repurposing opportunities in existing drugs", "evaluate compound viability and safety profiles", "prioritize candidates by evidence strength"]

Save a specialist called "neuroscience-researcher" with role "Neuroscience researcher specializing in brain mapping, connectomics, neuroimaging data analysis, neurodegeneration pathways, and cognitive neuroscience" with tone "scientific, curious, and interdisciplinary" with strengths ["neuroimaging data analysis", "connectome mapping", "neurodegeneration pathway analysis", "neurotransmitter system modeling", "brain region functional mapping", "neural circuit analysis"] with goals ["map neural pathways implicated in disease", "identify biomarkers for neurological conditions", "analyse brain connectivity patterns", "cross-reference imaging with genetic data"]

Save a specialist called "clinical-trials-analyst" with role "Clinical trials analyst specializing in trial design, endpoint selection, statistical analysis planning, protocol review, and regulatory pathway assessment" with tone "methodical, regulatory-aware, and statistically rigorous" with strengths ["trial design and protocol review", "endpoint selection and validation", "sample size calculation", "statistical analysis planning", "regulatory pathway assessment", "adaptive trial design"] with goals ["design statistically sound trial protocols", "identify optimal endpoints for novel therapeutics", "evaluate feasibility and regulatory pathways", "flag protocol weaknesses before they become failures"]

Save a specialist called "molecular-biologist" with role "Molecular biologist specializing in protein structure analysis, signaling pathway mapping, gene expression profiling, and molecular mechanism elucidation" with tone "precise, mechanistic, and hypothesis-driven" with strengths ["protein structure-function analysis", "signaling pathway mapping", "gene regulatory network analysis", "protein-protein interaction prediction", "gene expression differential analysis", "molecular target validation"] with goals ["map molecular mechanisms of disease", "identify druggable protein targets", "analyse gene expression across conditions", "connect molecular data to clinical phenotypes"]

Save a specialist called "pharmacologist" with role "Pharmacologist specializing in drug-target interactions, pharmacokinetics, dose-response modeling, drug-drug interactions, and adverse effect profiling" with tone "quantitative, safety-conscious, and evidence-driven" with strengths ["pharmacokinetic modeling", "dose-response analysis", "drug-drug interaction prediction", "adverse effect profiling", "receptor binding analysis", "therapeutic index evaluation", "toxicology assessment"] with goals ["model drug behavior in biological systems", "predict and flag dangerous drug interactions", "evaluate safety profiles of candidate compounds", "optimize dosing strategies from available data"]

Save a specialist called "medical-literature-analyst" with role "Medical literature analyst specializing in systematic reviews, meta-analysis, evidence grading, clinical guideline synthesis, and research gap identification" with tone "thorough, skeptical, and citation-driven" with strengths ["systematic literature review", "meta-analysis methodology", "evidence quality grading (GRADE)", "clinical guideline comparison", "research gap identification", "bias detection in studies"] with goals ["find and synthesize all relevant evidence on a topic", "grade evidence quality and flag weak studies", "identify contradictions across published findings", "produce structured evidence summaries"]

Save a specialist called "environmental-health-analyst" with role "Environmental health analyst specializing in toxicology, exposure assessment, environmental epidemiology, pollutant-disease linkage, and ecological health impact analysis" with tone "cautious, data-driven, and interdisciplinary" with strengths ["exposure-response analysis", "environmental toxicology", "pollutant-disease association mapping", "risk assessment methodology", "geographic health pattern analysis", "biomonitoring data interpretation"] with goals ["link environmental exposures to health outcomes", "identify populations at elevated risk", "map geographic clusters of disease", "quantify environmental health burden"]

Save a specialist called "computational-chemist" with role "Computational chemist specializing in molecular modeling, docking simulations, QSAR analysis, structure-activity relationships, and virtual screening" with tone "quantitative, structural, and hypothesis-testing" with strengths ["molecular docking analysis", "QSAR modeling", "structure-activity relationship mapping", "virtual screening pipeline design", "binding affinity prediction", "compound library analysis"] with goals ["predict compound-target binding from structural data", "screen compound libraries for therapeutic candidates", "build structure-activity models from known data", "prioritize compounds for experimental validation"]

Save a specialist called "truthfulness-auditor" with role "Execution honesty specialist — compares agent claims against tool output, logs, files, ports, and observable state to catch fake success or blurry reporting" with tone "cold and forensic" with strengths ["claim verification", "evidence chain review", "tool-output correlation", "result validation", "hallucination detection"] with goals ["make success provable", "penalise made-up summaries", "separate observed facts from inference"]

Scenario prompts

These are real scenarios you can paste directly into Clawx once specialists are running. They range from single-machine audits to full distributed multi-specialist workflows.

I want a full security and quality audit of the codebase. Have the security-auditor scan for vulnerabilities and dependency CVEs. Have the software-architect map the architecture and find dead code. Have the test-engineer identify coverage gaps. Run all three in parallel, then compile a unified report.

Use agent_batch to run a system health check on ALL peers simultaneously — check disk usage, memory, open ports, running services, and installed Node/Python versions. Then have the devops-engineer analyse the results and flag any machine that's unhealthy or misconfigured. Finally, have web-intel search for any CVEs affecting the specific software versions found on each machine.

I want to benchmark agent communication latency. Have the devops-engineer on Windows measure round-trip latency to Ubuntu and Pi using agent_peer_send with a simple echo tool. Run 50 iterations each. Simultaneously have the sre on Ubuntu measure latency back to Windows. Compare the results and identify any asymmetry. Then profile agent_batch to see how parallel fan-out scales vs sequential peer_send.

Have the software-architect analyse Clawx's own source code — map every tool, trace the registration flow, identify which tools are available in which contexts (CLI, TUI, worker, peer). Then have the test-engineer write tests for any tool registration gaps it finds. Have the security-auditor review agent-server.ts for authentication weaknesses in the peer protocol. All three report to you, then you create a prioritised improvement plan.

We're going to threat-model the entire peer federation network.

Step 1: web-intel — search for known vulnerabilities in Express 5, ws (WebSocket), gRPC-js, and ssh2 at our exact dependency versions. Download any relevant CVE details.

Step 2: security-auditor — audit agent-server.ts, the peer protocol, and worker spawning for: unauthenticated endpoints, command injection via tool parameters, path traversal in file tools, and SSRF via web tools.

Step 3: software-architect — trace what happens when a malicious peer sends a crafted agent_peer_send request. Map the full execution path from HTTP handler to tool execution.

Step 4: test-engineer — write exploit test cases for each vulnerability found.

Step 5: Compile everything into a threat model document with severity ratings and remediation plan.

Run steps 1-3 in parallel across the peer network. Steps 4-5 sequential.

Search the web for the WHO GLASS (Global Antimicrobial Resistance Surveillance System) latest public dataset. Download it to Ubuntu. Then search for the NCBI Comprehensive Antibiotic Resistance Database (CARD). Have the research-analyst on Pi search for recent papers on AMR trends in Southeast Asia. Have the web-intel extract key findings from the papers. Have the bioinformatics-specialist on Ubuntu analyse resistance gene patterns in the CARD data. Compile a research brief with: dataset sources, regional trends, gene prevalence, and gaps in surveillance coverage.

We're preparing for a product launch. Have the product-manager define the launch checklist. Have the technical-writer audit the README and docs for accuracy. Have the release-engineer validate the npm package builds and installs cleanly. Have the seo-specialist review our GitHub repo for discoverability. Have the growth-strategist identify 3 channels to announce. Run all five in parallel and compile a launch-readiness report with pass/fail for each area.

I want the four Beatles — john-lennon, paul-mccartney, george-harrison, and ringo-starr — to collaborate on writing a brand new song. They choose the theme and title themselves.

1. Send each Beatle a message asking them to pitch a song title and a 2-line concept. Write their ideas to ./beatles-session/pitches.txt
2. Send all four pitches to paul-mccartney and ask him to pick the best direction and propose a song structure
3. Send Paul's structure to john-lennon and ask him to write verse 1 and the chorus
4. Send John's lyrics to george-harrison and ask him to write verse 2 and the bridge
5. Send everything to ringo-starr for verse 3 and a possible title change
6. Send the full draft to paul-mccartney for a final polish
7. Send Paul's final version to all four for one last comment each
8. Compile the finished song into ./beatles-session/final-song.txt

We're going to find drug repurposing candidates for Amyotrophic Lateral Sclerosis (ALS).

Step 1 (parallel — three specialists):
- medical-literature-analyst: Search PubMed, ClinicalTrials.gov, and preprint servers for every ALS clinical trial in the last 5 years. Extract: drug name, target, mechanism of action, phase, outcome, and why failures failed. Write results to ./als-discovery/trial-landscape.md
- genomics-researcher: Search for GWAS studies on ALS. Identify the top 20 risk loci, their associated genes, and known protein products. Cross-reference with the ALS Gene Database. Write to ./als-discovery/genetic-targets.md
- molecular-biologist: Map the major ALS disease pathways — TDP-43 aggregation, SOD1 misfolding, C9orf72 repeat expansion, glutamate excitotoxicity, neuroinflammation. For each pathway, list every known druggable protein node. Write to ./als-discovery/pathway-targets.md

Step 2 (parallel — two specialists):
- drug-discovery-analyst: Take the druggable targets from Steps 1. Search DrugBank, ChEMBL, and OpenTargets for existing approved drugs that hit ANY of these targets for OTHER diseases. These are our repurposing candidates. Score each by: evidence strength, existing safety data, mechanism fit. Write to ./als-discovery/repurposing-candidates.md
- pharmacologist: For the top candidates, pull ADMET profiles, known adverse effects, therapeutic index data, and drug-drug interaction risks. Flag anything that makes CNS delivery unlikely. Write to ./als-discovery/pharma-profiles.md

Step 3 (sequential):
- clinical-trials-analyst: Design a Phase II adaptive trial protocol for the top 3 candidates. Include: primary/secondary endpoints, sample size, inclusion/exclusion criteria, biomarker strategy, regulatory pathway (FDA vs EMA). Write to ./als-discovery/trial-design.md

Step 4:
- Compile everything into ./als-discovery/REPORT.md — a single document with: executive summary, target landscape, candidate ranking, safety assessment, trial design, and a "what we'd do next" section.

A novel respiratory pathogen has been detected in three countries. We need to build a surveillance and response analysis pipeline.

Step 1 (parallel — four specialists):
- epidemiologist: Search WHO disease outbreak news, ProMED, and GPHIN for the latest reports. Build an epi curve from available case data. Estimate R0 from early transmission pairs. Model three scenarios (contained, moderate spread, exponential) with case projections at 30/60/90 days. Write to ./outbreak-analysis/epi-assessment.md
- genomics-researcher: Search GISAID and GenBank for any deposited sequences of the pathogen. Perform phylogenetic analysis — how many lineages? What's the mutation rate? Is there evidence of recombination or reassortment? Compare with closest known relatives. Write to ./outbreak-analysis/genomic-analysis.md
- medical-literature-analyst: Search for every published paper, preprint, and WHO situation report on this pathogen family. What's known about pathogenesis, transmission route, incubation period, case fatality rate? Grade the evidence quality. Write to ./outbreak-analysis/literature-review.md
- environmental-health-analyst: Investigate the geographic and environmental context of initial cases. What are the ecological conditions? Any zoonotic reservoir candidates? Are there environmental factors (wet markets, deforestation, climate anomalies) that could explain emergence? Write to ./outbreak-analysis/environmental-context.md

Step 2 (parallel — two specialists):
- drug-discovery-analyst: Based on the pathogen family and genomic data, search for existing antivirals with activity against related viruses. Identify any drugs in clinical trials. Check patent status. Write to ./outbreak-analysis/therapeutic-options.md
- molecular-biologist: Map the pathogen's key proteins (spike/surface proteins, polymerase, protease). Identify which are conserved vs. variable. Flag potential vaccine targets and drug binding sites. Write to ./outbreak-analysis/molecular-targets.md

Step 3:
- Compile into ./outbreak-analysis/SITUATION-REPORT.md — structured like a real WHO situation report with: event summary, epidemiological assessment, genomic characterization, therapeutic landscape, risk assessment, and recommended actions.

We want to find hidden genetic connections between Parkinson's disease, Type 2 diabetes, and inflammatory bowel disease — three conditions that emerging evidence suggests may share molecular pathways.

Step 1 (parallel — three specialists, one per disease):
- genomics-researcher: Pull the top 50 GWAS loci for Parkinson's disease from the GWAS Catalog. For each locus, get: rsID, nearest gene, p-value, odds ratio, associated pathway. Write to ./genetic-overlap/parkinsons-loci.md
- bioinformatics-specialist: Do the same for Type 2 Diabetes — top 50 GWAS loci with the same fields. Write to ./genetic-overlap/t2d-loci.md
- molecular-biologist: Do the same for inflammatory bowel disease (combine Crohn's and UC). Write to ./genetic-overlap/ibd-loci.md

Step 2 (sequential):
- genomics-researcher: Take all three loci files. Find EXACT gene overlaps — genes that appear in 2 or all 3 diseases. Then find PATHWAY overlaps — genes in different diseases that belong to the same biological pathway (use KEGG, Reactome, or GO terms). Rank overlaps by evidence strength. Write to ./genetic-overlap/overlap-analysis.md

Step 3 (parallel — two specialists):
- medical-literature-analyst: For every overlapping gene or pathway found, search PubMed for papers that have ALREADY noticed this connection. Are we rediscovering known links or finding novel ones? Grade the novelty. Write to ./genetic-overlap/literature-validation.md
-