AgentKit Preview

Next-gen AI coding orchestration — AgentKit Swarm, Graphify, and live auto-update

🚨 PREVIEW BUILD — This is the cutting-edge preview of AgentKit. For the stable public build, see agentkit-ai.

🚀 Preview v0.5.65 — AgentKit Swarm, Graphify 3D, live auto-update, OpenCode MCP watcher.

agentkit-preview vs agentkit-ai

| Package | Version | Features | Install | |---------|---------|----------|---------| | agentkit-preview (here) | 0.5.64 | Swarm + Graphify + Live auto-update + MCP watcher | npx agentkit-preview init | | agentkit-ai (stable) | 0.5.63 | Core skills + memory + workflow | npx agentkit init |

AgentKit vs Standalone Graphify — Full Benchmark

AgentKit ships its own code graph engine. No separate install. No MCP server. Wins on every dimension:

| Feature | Standalone Graphify | AgentKit Built-in | Result | |---------|--------------------|--------------------|--------| | Installation | Manual install / MCP server | Zero-install — built into agentkit init | ✅ WIN | | Languages | Multi-language | 26 languages (Python, Go, Rust, Java, TS, C++, Swift, Kotlin…) | ✅ WIN | | Parse accuracy | tree-sitter AST | Tree-sitter AST (complete, regex fallback) | ✅ WIN | | CLI | Standalone binary | agentkit graphify build/watch/search/nl/stats/diff/export | ✅ WIN | | Visualization | Canvas/SVG via vis-network | 3D interactive — glossy spheres, depth fog, orbit controls, inspector panel | ✅ WIN | | Search | Basic keyword | Natural Language Search — query in plain English | ✅ WIN | | Integrations | MCP / External only | Watchdog + Memory + Context Injector (every turn) | ✅ WIN | | Intelligence | Call graph + communities | Communities + Blast Radius + NL + Live topology injection | ✅ WIN |

3D Code Graph Visualization

AgentKit generates a self-contained 3D interactive graph of your entire codebase at graphify-out/graph.html.

Two ways to open it:

# Option 1 — CLI
agentkit graphify visualize

# Option 2 — direct (works anywhere, no command needed)
open graphify-out/graph.html        # macOS
xdg-open graphify-out/graph.html   # Linux
start graphify-out/graph.html      # Windows

What you see:

• 25,715+ nodes — every function, class, method, and module in your repo
• Glossy 3D spheres with depth fog and perspective projection — rendered at 60fps via Canvas2D
• Color-coded by type: emerald = modules · blue = classes · gold = functions · lavender = methods · amber = god nodes
• God nodes (high-connectivity hubs) rendered larger — the files that break everything when touched
• Right-side inspector — click any node to see its degree, cluster, blast radius, and all connected neighbors
• Blast radius on click — 2-hop BFS highlights every node that would be affected by a change
• Search + type filters on the left — find any symbol in milliseconds
• Auto-rotate — orbits the graph until you interact; drag to orbit, scroll to zoom, shift+drag to pan

How it flows into the model:

agentkit init
  └─ builds graphify-out/graph.json  (25,715 nodes · 1,943 edges · 52 god nodes)

Every prompt →  memory/injector.py
  ├─ SQLite memory      (1,400 tokens — decisions, entities, past sessions)
  └─ Code graph         (1,400 tokens — callers, callees, blast radius, god node warnings)

The model always knows what it's about to touch, what calls it, and what breaks if it changes — before writing a single line.

What's new in v0.5.59 — Auto-Incremental Graph Build

One problem fixed:

The code graph was static — built once during agentkit init and frozen until manually rebuilt. If you added 50 new files or renamed functions, the graph still showed the old state.

Solution: Added a PostToolUse hook that triggers incremental graph rebuilds after every file edit.

Claude edits a file
  └─ PostToolUse hook → agentkit graphify build (incremental, background)
       └─ only re-parses changed file, ~500ms
       └─ updates graph.json + graph.html

What's included

| Feature | Description | |---------|-------------| | Auto-incremental build | After every Edit/Write/MultiEdit, the graph updates automatically | | Fast rebuilds | Uses mtime caching — only re-parses changed files (~500ms vs ~60s) | | Background execution | Runs async, doesn't block the model | | Live topology injection | Every turn gets updated callers, callees, blast radius |

Manual options (still available)

agentkit graphify build    # One-shot rebuild
agentkit graphify watch    # Live watch mode (run in separate terminal)

Automatic Watchdog Mode

AgentKit ships its own graph engine. No external dependency. On every agentkit init:

agentkit init
  └─ auto-builds graphify-out/graph.json
       25,715 nodes · 1,943 edges · 52 god nodes

Every prompt
  ├─ SQLite memory   (1,400 tokens — decisions, entities, past sessions)
  └─ Code graph      (1,400 tokens — callers, callees, blast radius, god nodes)

What the model sees on every turn:

## Code Graph Context
**`memory/injector.py`** — 3 symbols, blast radius: 262 nodes
  ⚠ God nodes in blast radius: parse, get, set +5 more
  - `build_injection` (function)  ← called by: run_router, main
  - `build_graph_injection` (function)  → calls: KnowledgeGraph, get_context_for_task

Automatic Watchdog Mode

AgentKit now automatically detects and fixes errors after every failed bash command — no manual invocation needed:

$ npm run dev
Error: npm ERR! code E404
[AgentKit] ✦ Auto-Watchdog triggered
[AgentKit] → Detecting error signature... module_not_found
[AgentKit] → Attempting fix: npm install requests
[AgentKit] ✓ Fix applied (attempt 1/3)
[AgentKit] Re-running: npm run dev

What's fixed automatically:

• Missing Python dependencies → pip install <package>
• Missing Node dependencies → npm install <package>
• Port conflicts → redeploy or fix Procfile
• Health check timeouts → bump gunicorn timeout

Risk tiers: | Tier | Behavior | Requires | |------|----------|----------| | Safe (deps, port) | Auto-apply | — | | Risky (git push) | Telegram approval | HITL | | Unknown | Terminal state | Manual review |

Blast Radius Detection

The killer feature: when something breaks, AgentKit knows what's affected:

from watchdog.graphify_enhancer import get_blast_radius

# On any failure, get the blast radius:
result = get_blast_radius("AttributeError: module 'x' has no attribute 'y'")
# → {
#   'failed_function': 'x.y',
#   'callers': ['a.py', 'b.py'],        # what uses it
#   'tests': ['test_x.py', 'test_y.py'],  # what covers it
#   'fix_suggestions': [...]
# }

This makes fixes context-aware — not just "what failed" but "what depends on it."

$ npx agentkit-preview@latest init

AgentKit Preview Installer v0.5.65
──────────────────────────────────

Detecting platforms...
  ✓ Claude Code  (full)
  ✓ OpenCode     (full)
  ✓ Cursor       (partial)

Using skill bundle: Backend Pro (22 skills)

Installing for Claude Code...
  ✓ Skills converted (SKILL.md native)
  ✓ Hooks registered → ~/.claude/settings.json
  ✓ Model routing: Haiku / Sonnet / Opus

Building code graph for memory injection...
  ✓ Code graph: 25974 nodes, 2383 edges, 64 god nodes, 154 communities

──────────────────────────────────────────────────
✓ AgentKit Preview installed successfully!

  Hooks registered in: ~/.claude/settings.json (global)
  AgentKit Preview will activate automatically in every Claude Code session.

  Estimated savings:
    Tokens:  ~40,000 → ~5,000/session  (89% reduction)
    Cost:    ~70% reduction vs default all-Sonnet

Demo GIF coming soon — record yours and open a PR!

How to Test the Preview

# Clone the preview repo
git clone https://github.com/Ajaysable123/AgentKit-Preview-01.git
cd AgentKit-Preview-01

# Install dependencies
pip install -r requirements.txt

# Run the test suite
pytest tests/test_graphify_integration.py -v

# Or run a smoke test
python -m pytest tests/test_watchdog_providers_smoke.py -v

Expected results:

• Graphify integration: 11/11 tests pass
• Watchdog: 58/58 provider tests pass
• Classifier: 93/93 classification tests pass
• Full integration: all smoke tests pass

Before vs After

Real numbers from AgentKit smoke tests, measured across a 50-turn coding session.

| Metric | Without AgentKit | With AgentKit | Improvement | |--------|-----------------|---------------|-------------| | Tokens per session | 45,000 | ~5,000 | 89% less | | Cost per session (Sonnet) | ~$1.35 | ~$0.40 | 70% cheaper | | Skill activation rate | 20% (ad-hoc) | 84% (hook-enforced) | 4× more reliable | | Model used for simple tasks | Sonnet ($0.003/K) | Haiku ($0.00025/K) | 12× cheaper | | Model used for subagents | Sonnet | Haiku (always) | 12× cheaper | | Context at session start | Full 10K token dump | 2K relevant nodes | 80% less noise | | Memory across sessions | None | SQLite graph + handoff | Persistent | | Coding without a plan | Allowed | Blocked by hook | Zero skipped steps |

One Command Install

npx agentkit-preview@latest init

That's it. AgentKit Preview detects your platforms, installs the right skill format for each, wires all hooks, and configures model routing automatically.

Or install globally (then use agentkit as a command anywhere):

npm install -g agentkit-preview
agentkit init

Note: The npm package name is agentkit-preview. After a global install, the CLI command is agentkit.

Requirements: Node.js ≥ 18 · Python ≥ 3.9 · Claude Code (for full feature set)

AgentKit Swarm — Multi-Agent Orchestration

AgentKit Preview includes AgentKit Swarm, a model-agnostic multi-agent orchestration framework that beats Ruflo while leveraging built-in skills, memory, and workflow enforcement.

$ agentkit swarm create "Build a REST API with auth, tests, and CI/CD"

[AgentKit Swarm] Orchestrator analyzing task...
[AgentKit Swarm] Task decomposed: architect → writer → tester → devops
[AgentKit Swarm] Starting 5-agent DAG with Byzantine consensus...

Wave 1 →  architect     Design schema + endpoints
Wave 2 →  writer         Implement API + auth
Wave 3 →  tester         Write pytest suite
          security       OWASP audit
Wave 4 →  devops         CI/CD pipeline
Wave 5 →  reviewer       Final code review

[AgentKit Swarm] All agents converged. 2/3 Byzantine consensus reached.

Competitive advantages: | Feature | Ruflo | AgentKit Swarm Preview | |---------|-------|------------------------| | Model Support | Claude only | ANY (OpenAI, Anthropic, Ollama, MiniMax) | | Skills | Basic prompts | 50+ built-in skills | | Memory | External setup | Built-in Graphify | | Workflow | Optional | Enforced R→P→E→R→S | | Cost | Sonnet for all | Free models (Haiku, Ollama) for workers |

Key modules:

• swarm/core/ — ModelPool, CostBudget, Topology, Agent, Task, Message
• swarm/adapters/ — OpenAI, Anthropic, Ollama, MiniMax, Factory
• swarm/orchestration/ — Planner, Coordinator, Monitor
• swarm/consensus/ — Byzantine BFT, Majority, Weighted voting
• swarm/roles/ — 10 pre-defined roles (researcher, architect, writer, tester, reviewer, security, devops, database, docs, integration)
• swarm/cli.py — swarm create/run/list-models/status

What It Does

AgentKit Preview is a 6-layer runtime that sits between your prompts and the model:

• Layer 0 — AgentKit Swarm: Model-agnostic multi-agent orchestration with Byzantine/Majority/Weighted consensus, 10 pre-defined roles, and DAG-based task execution
• Layer 1 — Skill Router: Classifies every prompt in < 10ms → loads only relevant skills → 45,000 tokens/session down to 5,000 (89% reduction)
• Layer 2 — Memory Graph: SQLite knowledge graph captures files, functions, decisions across sessions → Haiku-compressed handoffs so context survives restarts
• Layer 3 — Graphify: 3D interactive code knowledge graph with 26 languages, community detection, god nodes, blast radius, natural language search, and live auto-update
• Layer 4 — Token Budget: Auto-routes Haiku / Sonnet / Opus by task complexity + proactive context compaction at 60% fill + real-time cost dashboard in your status bar
• Layer 5 — Workflow Engine: Enforces Research → Plan → Execute → Review → Ship via hooks — can't skip planning, quality gates (syntax/lint/types/tests) run after every edit
• Layer 6 — Platform Layer: One SKILL.md file auto-converted to 10 platform formats — Cursor .mdc, Codex AGENTS.md, Gemini CLI config, and more

Works With

| Platform | Support | Install format | |----------|---------|----------------| | | Full — skills + hooks + memory + routing | Native SKILL.md | | | Skills + model routing rules | .cursor/rules/*.mdc | | | Skills via system prompt | .gemini/GEMINI.md | | | Skills via Cascade rules | .windsurf/rules.md | | | Skills + native TUI plugin + slash commands | System prompt + TUI plugin | | | Skills as plugins | .kilo/plugins/*.yaml | | | Skills injected | AGENTS.md | | | Skills as conventions | .aider.conf.yml | | | Skills as context | .augment/context.md | | | Full plugin system | .antigravity/plugins/ |

Ruflo: AgentKit Preview makes your Ruflo swarms 3× cheaper by routing worker agents to Haiku and injecting only relevant skills per agent.

OpenCode Integration

AgentKit ships a native TUI plugin for OpenCode that lives inside the terminal UI — not just in the system prompt.

What you get

| Feature | Detail | |---------|--------| | AgentKit agent persona | agentkit ⚡ appears in OpenCode's agent list — press tab to switch to it | | Active agent label | Status bar shows Agentkit in orange when selected | | Startup toast | ⚡ AgentKit v0.5.x Active — 54 skills loaded appears on every launch | | /agentkit | Status command — shows version, skill count, session cost | | /agentkit-task | Pre-fills the prompt with @agentkit-task: — type your task and press Enter | | /agentkit-analytics | Shows cost & usage info | | /ak | Alias for /agentkit | | /ak-task | Alias for /agentkit-task |

The AgentKit Agent Persona

After install, AgentKit registers itself as a named agent in OpenCode's agent switcher. Press tab → select agentkit ⚡ to activate it. Once active, the status bar shows Agentkit in orange — you're now running with full AgentKit skill-routing and workflow enforcement on every message.

Install (new users)

npx agentkit-preview@latest init

AgentKit Preview auto-detects OpenCode and installs everything — system prompt injection, TUI plugin, slash commands, MCP servers, and the agent persona. Restart OpenCode after running init.

Update (existing AgentKit Preview users)

npm install -g agentkit-preview@latest
agentkit init

AgentKit auto-detects OpenCode and installs everything — system prompt injection, TUI plugin, slash commands, and the agent persona. Restart OpenCode after running init.

Update (existing AgentKit users)

npm install -g agentkit-ai@latest
agentkit init

Then restart OpenCode. If you installed before v0.5.18, this adds the agent persona to the tab → agents switcher.

How to assign a task

Option A — via agent + prompt (recommended):

1. Press tab → select agentkit ⚡
2. Type your task directly and press Enter

Option B — via slash command:

1. Press ctrl+p → type /agentkit-task → select it
2. The prompt pre-fills with @agentkit-task:
3. Type your task after the prefix and press Enter

Both options activate the full AgentKit workflow: Research → Plan → Execute → Review → Ship.

Manual plugin install (optional)

If auto-detection misses OpenCode, register the plugin directly:

opencode plugin "/path/to/AgentKit/platform/opencode-plugin" --global --force

Find the path with: npm root -g agentkit-preview → append /platform/opencode-plugin.

Dynamic Agent Spawning

AgentKit v0.5.4 adds a zero-config spawn engine that automatically detects when a task needs multiple agents and orchestrates them for you.

$ claude "Build a REST API with auth, tests, and a security audit"

[AgentKit] Multi-agent task detected (confidence: 0.90)
[AgentKit] Spawning 5 agents in 4 waves...

Wave 1 →  architect  (opus-4.6)      Design schema + endpoints
Wave 2 →  writer     (haiku-4.5)     Implement API + auth          [waits: architect]
Wave 3 →  tester     (haiku-4.5)     Write pytest suite            [waits: writer]
         security   (sonnet-4.6)    OWASP audit                   [waits: writer]
Wave 4 →  reviewer   (sonnet-4.6)    Final code review             [waits: writer + tester]

• 3-tier detection: keyword signals (<5ms) → heuristic scoring (<10ms) → Haiku LLM fallback (~$0.0003) for ambiguous cases
• Smart model routing per role: Architect gets Opus, implementation gets Haiku, security/review get Sonnet
• DAG execution: parallel where possible, sequential where dependencies require it
• Recursion-safe: spawned agents never re-spawn (infinite loop prevention built-in)

How AgentKit Compares

| Feature | AgentKit | Superpowers | claude-mem | ClaudeFast | |---------|----------|-------------|------------|------------| | Dynamic agent spawning | ✅ Auto-detects, N agents, DAG waves | ❌ | ❌ | ❌ | | Smart skill loading | ✅ Auto-routed, 89% token reduction | ✅ Manual SKILL.md | ❌ | ❌ | | Skill library | ✅ 50+ skills, 7 role bundles | ❌ BYO only | ❌ | ❌ | | Persistent memory | ✅ SQLite graph + session handoffs | ❌ | ✅ Basic | ❌ | | Auto model routing | ✅ Haiku/Sonnet/Opus by complexity | ❌ | ❌ | ⚠️ Basic | | Workflow enforcement | ✅ Research→Plan→Execute→Review→Ship | ⚠️ Suggested only | ❌ | ❌ | | Quality gates | ✅ syntax+lint+types+tests on every edit | ❌ | ❌ | ❌ | | Multi-platform | ✅ 10 platforms, 1 config | ❌ Claude Code only | ❌ | ❌ | | Subagent cost routing | ✅ Per-role model (12× cheaper) | ❌ | ❌ | ❌ | | Cost dashboard | ✅ Real-time status bar | ❌ | ❌ | ✅ | | npx install | ✅ One command | ❌ Manual | ❌ Manual | ❌ |

CLI Reference

Without global install (use npx agentkit-preview <command>):

npx agentkit-preview@latest init      # First install / update
npx agentkit-preview sync             # Re-sync after adding skills
npx agentkit-preview status           # Health check + cost summary
npx agentkit-preview analytics        # Cost & usage dashboard

With global install (npm install -g agentkit-preview, then use agentkit):

agentkit init              # Detect platforms → install
agentkit sync              # Re-sync after adding skills
agentkit status            # Health check + cost summary
agentkit costs --days 7     # Weekly cost analytics
agentkit skills list        # Browse all 50+ skills
agentkit workflow status    # Current Research/Plan/Execute state
agentkit workflow approve   # Approve plan → unlock coding
agentkit detect             # Show detected AI coding tools
agentkit uninstall          # Remove all AgentKit files
agentkit uninstall --purge  # Also delete runtime data (costs/memory/state)

With global install (npm install -g agentkit-ai, then use agentkit):

agentkit init              # Detect platforms → install
agentkit sync              # Re-sync after adding skills
agentkit status            # Health check + cost summary
agentkit costs --days 7    # Weekly cost analytics
agentkit skills list       # Browse all 50 skills
agentkit workflow status   # Current Research/Plan/Execute state
agentkit workflow approve  # Approve plan → unlock coding
agentkit detect            # Show detected AI coding tools
agentkit uninstall         # Remove all AgentKit files
agentkit uninstall --purge # Also delete runtime data (costs/memory/state)

Skill Bundles

Pick a bundle at install time or pass --bundle <name>:

| Bundle | Skills | Best for | |--------|--------|----------| | backend-pro | python-debugger, go-debugger, pytest, rest-api, grpc, sql, mongodb, redis, auth, owasp, docker, nginx + 10 more | Python/Go backend engineers | | frontend-wizard | js-debugger, jest, cypress, playwright, react, vue, nextjs, css, state-mgmt, a11y, graphql + 2 more | Frontend / React developers | | full-stack-hero | All 50 skills | Full-stack teams | | ai-engineer | llm-prompting, rag-pipeline, function-calling, agent-design, eval-testing + 5 more | LLM / AI application developers | | devops-master | docker, kubernetes, github-actions, terraform, monitoring, nginx + 3 more | DevOps / Platform engineers | | data-scientist | pandas, data-viz, ml-pipeline, sql, pytest + 2 more | Data scientists / ML engineers | | mobile-dev | react-native, flutter, rest-api, auth-jwt + 3 more | Mobile developers |

All 50 Skills

| Category | Skills | |----------|--------| | Debugging | python-debugger, js-debugger, go-debugger, network-debugger | | Testing | tdd-workflow, jest-testing, pytest-workflow, cypress-e2e, playwright-testing, contract-testing | | API | rest-api, graphql, grpc, openapi-design, webhook-design | | Database | sql-query, prisma-orm, mongodb, redis-caching, database-migrations | | Frontend | react-patterns, nextjs-patterns, css-layout, vue-patterns, state-management, accessibility | | DevOps | docker, kubernetes, github-actions, terraform, monitoring-observability, nginx-config | | Security | auth-jwt, owasp-top10, secrets-management, api-security | | Refactoring | clean-code, performance-optimization, code-review, legacy-modernization | | AI Engineering | llm-prompting, rag-pipeline, function-calling, agent-design, eval-testing | | Data Science | pandas-workflow, data-visualization, ml-pipeline | | Mobile | react-native, flutter |

Built on the shoulders of giants: Superpowers (108K ⭐) · claude-mem (39.9K ⭐) · awesome-claude-code (30.9K ⭐)

npm · GitHub · Issues · MIT License