agent-compiler

Agents are compiled, not stored. Humans author small, addressable behavior modules; a model may turn fuzzy intent into a typed AgentQuery, but everything after that boundary is deterministic: the kernel resolves modules, expands dependencies, fails closed on conflicts and over-ceiling effects, and emits an immutable, content-hashed AgentImage that renders into a harness-native agent definition — every line traceable to its source module.

Install

/plugin marketplace add JRichlen/agent-plugins
/plugin install agent-compiler@jrichlen

Try it

cd plugins/agent-compiler
python3 scripts/compile.py compile \
  --registry registry \
  --query examples/queries/security-pr-review.json \
  --out /tmp/image.json
python3 scripts/render_claude_agent.py --image /tmp/image.json

Compile it twice — the bytes and the hash are identical. Add an unrelated module — the hash does not move. Ask for a write capability under a read-only ceiling — compilation fails with EFFECT_CEILING instead of shipping a quietly over-privileged agent.

Installed as a Claude Code plugin, the bundled MCP server starts automatically and exposes the same operations as tools — inspect, compile, explain, render — read-only by construction (no execute, ever). The CLI above is the identical surface for any harness without MCP.

Pieces

• skills/agent-compiler/SKILL.md — the boundary discipline: natural language may select behavior, never silently define it.
• skills/agent-compiler/references/language.md — the module format and exact selection semantics.
• scripts/compile.py — the deterministic kernel (stdlib-only python3): compile, inspect, explain.
• scripts/render_claude_agent.py — AgentImage → Claude Code agent definition (the image is the contract; other harnesses get sibling renderers).
• scripts/mcp_server.py — the auto-started MCP facade (stdlib-only JSON-RPC over stdio), declared under mcpServers in .claude-plugin/plugin.json.
• hooks/ — two reinforcement hooks (Claude Code only; a convenience, not a dependency — the skill carries the same rules on any other harness): a prompt-time nudge toward the MCP tools on agent-building intent, and a guard that denies hand-edits of compiled artifacts.
• registry/ — starter modules; --registry accepts any directory, so your repo can carry its own.
• evals/cheap/ — golden, metamorphic, and fail-closed checks run by the marketplace cheap tier.

Design history: docs/designs/agent-compiler-plugin.md at the repo root.