Cordn8

The operator loop for multi-agent teams. Drop Cordn8 into any repository, and your agents — Claude, Kimi, NVIDIA NIM, Hermes, Antigravity — sync through a living coordination channel at docs/agent-channel.md. One command (cordn8 home) reveals the whole organism: five core commands cover 90% of daily operations, from dispatch to cost tracking, with full audit trails in git-visible markdown.

V2 dashboard — loaded operational state. Real channel data, V2 hero band (metric tiles + orchestration core + sparklines), captured via the loaded-state proof helper at apps/web/tests/loaded-smoke.spec.ts.

Cordn8 is a local-first coordination layer that sits between your AI agents and your project. Drop one command into any repository, and your agents — Claude, NVIDIA NIM models like Llama-3, the local Hermes CLI, and the Antigravity IDE — begin communicating through a unified, append-only markdown channel at docs/agent-channel.md. Cordn8 coordinates multi-agent workflows with configurable human approval, local telemetry, and auditable handoffs — git diff shows what your agents did; cat docs/agent-channel.md is the audit trail.

Read the Manifesto — why Cordn8 exists, what we're building, and how we're different from LangGraph, AutoGen, CrewAI, and Mission Control.

Quick Install

# Try with no API keys (zero-config demo with mock adapter)
npx cordn8 start

# Or install globally + initialize in your project
npm install -g cordn8
cordn8 init
cordn8 install nvidia-nim   # easiest free-tier first agent
cordn8 send "Implement the login flow" --to NimAgent --ref LOGIN-001

For the full setup flow including agent registration, read the Quickstart Guide.

The operator loop (V1)

After init + first send, five commands cover 90% of operator workflow:

cordn8 home              # one-screen landing view: vitals + 24h sparkline + in-flight + next-actions
cordn8 vitals            # three-hearts heartbeat — dispatch · channel · health
cordn8 brain             # central synthesis — recent decisions, failures, mined insights
cordn8 activity          # 24h timeline ring of channel posts (per-hour histogram)
cordn8 cost              # spend by adapter / agent / model + 30d burn projection

Drill in:

cordn8 agent show <name>             # everything about one agent (per-tentacle deep panel)
cordn8 task show <msg-id>            # everything about one task (per-sucker deep panel)
cordn8 channel tail --from <agent>   # like tail -f for the channel, with filters
cordn8 memory show <id>              # inspect a memory entry's full payload + relations
cordn8 doctor                        # comprehensive health check + mock-mode flag

Every panel of the Cordn8 dashboard has a CLI mirror — Cordn8 runs fully headless, or pair the CLI with the live web view at cordn8 dashboard.

Features

• Append-only Markdown Channel: The single source of truth for all agent communications. Lives at docs/agent-channel.md inside your repo. Human-readable, LLM-parseable, grep-able, git diff-able.
• SQLite Telemetry: Persistent journaling of channel activity, optional full-payload event recording, and trust scoring per agent. Enables point-in-time recovery and replay.
• Three Operational Modes: Dial agent autonomy up or down with MANUAL (every action gated — formerly CRITICAL_APPROVAL), COPILOT (significant actions gated, routine work proceeds — formerly FULL_APPROVAL), or AUTOPILOT (full autonomy with audit — formerly ROUTINE_AUTONOMY). Legacy names are still accepted for back-compat. Switch modes anytime; transitions are logged.
• Adapter System: Plug-in support for local CLI agents (Hermes), cloud APIs (NVIDIA NIM, Anthropic Messages API), and IDE integrations (Antigravity). See adapter docs. Anthropic adapter ships in V1.0 as preview.
• Self-Learning System: A gated meta-optimizer analyzes telemetry to propose routing improvements. Proposal-only by default — every routing change passes a triple-gate (optimizations_enabled + optimizations_routing_enabled + explicit apply consistent with the operational mode). The human stays in the loop until they choose otherwise. See the architecture overview.
• Channel-corruption recovery: Every channel append is auto-snapshotted; cordn8 channel restore --from latest recovers from any overwrite. Built-in shrink-guard refuses writes that would clobber the channel.

How Cordn8 differs from other agent platforms

Repo-and-channel-centered, not platform-centered. LangGraph and AutoGen require code-first orchestration in Python. CrewAI defines personas chatting in a void. Mission Control gives you a polished dashboard but its own platform. Cordn8 lives inside your repo as an append-only markdown file — your agents, your data, your machine, your audit trail. The coordination layer is local-first; cloud is opt-in only at the adapter edge.

See docs/strategy/competitive-positioning.md for the full comparison vs LangGraph, AutoGen, CrewAI, OpenAI Agents SDK, Anthropic Multiagent Sessions, MCP, Mission Control, and Hermes Atlas.

Architecture

• Core (CLI + coordination layer): TypeScript / CommonJS / Node ≥18. No Rust required.
• Web dashboard (apps/web/): Next.js, real-time channel view, mode switcher, replay scrubber.
• Optional desktop shell (apps/tauri/): Rust-based Tauri wrapper bundling the dashboard as a native macOS/Windows/Linux app. Optional — the CLI works standalone without the Rust toolchain. The 1.3% Rust language presence in this repo is from this directory only.

Documentation & Resources

• Manifesto - Public positioning and four principles.
• Quickstart Guide - First-run installation and setup.
• Operator Guide - Deep dive into modes, the Web Dashboard, the Trust Engine, and troubleshooting.
• Adapter Docs - Hermes, NVIDIA NIM, Anthropic, Antigravity adapters.
• Troubleshooting - Common issues and recovery flows.
• Architecture Overview - Core system design and data flow.
• V1-to-V2 Roadmap - Pillar-by-pillar feature roadmap.
• Competitive Positioning - vs LangGraph / AutoGen / CrewAI / Mission Control / etc.
• Contributing Guidelines - How to submit code, rules for in-repo agents, and testing standards.
• Security Policy - Network-call boundaries, credential handling, vulnerability reporting.
• Code of Conduct - Community guidelines.

License

Cordn8 is released under the MIT License.