Skills for AI coding assistants
An MIT-licensed, graph-native skill runtime with zero third-party skill content and every skill quality-gated on every commit.
Watch it render
A 9-second tour of Agent Foundry, produced by HyperFrames (Apache-2.0, 36k★) from HTML to MP4 with FFmpeg + Puppeteer. hyperframes render . --quality draft
One command, one daemon
Agent Foundry runs entirely locally. Install, run, verify, no cloud, no SaaS, no telemetry.
▸ Installing agent-foundry for claude-code...
✔ Installed: ~/.claude/skills/<39 skills>
✔ 39 skills · 3 agents · 22 CLI commands
$ agent-foundry doctor
✔ Config: ~/.config/agent-foundry/config.toml
✔ Index: 39 skills loaded
✔ Daemon: running on 127.0.0.1:8765
✔ API key: ANTHROPIC_API_KEY configured
$ agent-foundry plan "refactor auth middleware"
1. bottleneck-gating (score 0.82)
2. verify-first (score 0.70)
$ agent-foundry run "refactor auth middleware"
→ Using bottleneck-gating...
gears 39 curated skills
Each ≤150 lines with trigger phrases, anti-patterns, and verification checklists. Disciplines, not domain knowledge.
brain 3 agents
af-planner decomposes requests. af-critic scores output. af-orchestrator dispatches subtasks to agents and skills.
graph Graph-aware planning
911 nodes, 1,860 edges, 71 communities. The knowledge graph feeds back into the planner, boosting related skills 15% higher in the ranking.
refresh Self-improvement loop
Rate outputs, run agent-foundry learn, and the system suggests new trigger patterns. Instincts are user-approvable and MIT.
What you get
Agent Foundry upgrades your workflow without forcing a tool reset.
Everything in one tool
Skills, agents, orchestrator, graph, self-improvement, built to work together.
Skills
Disciplines the model applies to its own work, how to think, not what to know. Each ≤150 lines with trigger phrase, anti-patterns, and verification. 25 core + 10 optional.
Agents
Roles the orchestrator dispatches. af-planner, af-critic, and the new af-orchestrator which decomposes → dispatches → merges subtasks across skills and agents.
Orchestrator
Local daemon: ranks, dispatches, executes, logs, judges. FastAPI, SQLite, lazy-started. HTTP endpoints for everything: /plan, /execute, /loop, /health.
Engram routing
Adopted from DeepSeek's Engram paper. O(1) N-gram hash → SQLite lookup for hot prompts. Plan + execute the same task in microseconds when the pattern is recognized.
Self-improvement
Rate executions, run agent-foundry learn, and the system suggests new trigger patterns from successful runs. Instincts stored in SQLite, user-approvable, MIT.
Knowledge graph
911 nodes, 1,860 edges, 71 communities. Tree-sitter AST extraction, community detection, god nodes, surprising connections. Graph-aware planning boosts related skills 15%.
Quality gates
Three gates on every commit: 39/39 validate, dependency-free MCP server, version drift check. Every skill is original, verified, and MIT-licensed.
14 harness adapters
Claude Code, Codex, Cursor, Hermes, Gemini, OpenCode, Antigravity, CodeBuddy, JoyCode, Qwen, Kimi, OpenClaw, Claude project, and Zed. Every target passes the installer + indexer smoke test.
22 commands, one local runtime
Every command is available from the CLI. The daemon starts lazily when a command needs it.
agent-foundry init
Write config and the first skill index. --force rewrites both.
agent-foundry index
Rebuild the skill index from disk.
agent-foundry plan "<prompt>"
Rank matching skills with graph boost, learned weights, and optional rerank.
agent-foundry run "<prompt>"
Plan, execute the top skill, optionally judge, and log the result.
agent-foundry execute <skill> "<prompt>"
Run one named skill directly.
agent-foundry resume <id> "<prompt>"
Continue a stored loop session with full history replay.
agent-foundry session <id>
Inspect a loop session, its turns, output, and status.
agent-foundry doctor
Health-check config, index, daemon, and API key.
agent-foundry status
Routing accuracy, fallback rate, and average cost.
agent-foundry consult "<need>"
Recommend skills for an unmet need.
agent-foundry cost-report
Token and time estimates per skill.
agent-foundry serve
Start the FastAPI daemon on 127.0.0.1:8765.
agent-foundry feedback <id> <score>
Attach a human score to an execution.
agent-foundry learn
Learn trigger weights from judged executions.
agent-foundry instincts
List proposed trigger instincts.
agent-foundry instinct-approve <id>
Approve an instinct and persist it.
agent-foundry mcp
Start the stdio MCP server: plan, execute, list_skills, search_skills, get_skill, verify_skills, check_action, runtime_status, execution_history, list_models.
agent-foundry browse
Browse or search the catalog; --read <id> prints a skill.
agent-foundry eval --tasks <file>
Paired evaluation with and without a skill, reporting lift.
agent-foundry attest
Sign a manifest of skill file hashes and frontmatter digests.
agent-foundry verify
Verify the installed manifest and signature; --strict exits on any mismatch.
agent-foundry gate
Classify a shell, file, text, or network action as allow, confirm, or deny.
By the numbers
Live from the repo, regenerated on every build, no hand-maintained counts.
Harness adapters
Every target is tested by the installer + indexer smoke test and ships 39 skills as direct children.
Featured skills
A taste of the catalog. See all 39 →
FAQ
Quick answers to common questions.
Agent Foundry ships 39 original skills with a hard zero-external-references gate, every skill is original MIT-licensed work. We also ship a live knowledge graph of the codebase and a 3-gate audit pipeline on every commit, including an external-reference scanner that fails the build if any third-party name leaks in.
Every commit runs a tree-sitter AST extraction of the entire codebase. The resulting graph (911 nodes, 1,860 edges, 71 communities) feeds back into the orchestrator, related skills get a 15% boost in planning. The graph is a first-class feature, not a visualization afterthought.
Yes. MIT-licensed, all local, no cloud. You supply the API key for your LLM provider. No SaaS, no telemetry, no subscription.
All 14 adapters are tested: Claude Code, Codex, Cursor, Hermes, Gemini, OpenCode, Antigravity, CodeBuddy, JoyCode, Qwen, Kimi, OpenClaw, Claude project, and Zed. Install one with node scripts/install.js --harness=NAME or use the npm, PyPI, or MCP package.
Absolutely. Fork the repo, add a skill under skills/core/ following the template in community/README.md, and open a PR. Every skill must pass foundry-eval.py and the nox.sh external-reference gate.
The daemon is lazy-started on first command. No systemd, no launchd, no Docker. Everything runs locally in ~/.config/agent-foundry/executions.db.
Ship with confidence
39 curated skills, 3 agents, a live knowledge graph, a self-improvement loop, and three quality gates, all MIT, all local.