Documentation

Install the skills, optionally install the runtime, then verify what your harness can see. Version 0.6.0.

Install skills

Node.js 22.14 or newer is required for the installer lane. The installer detects a harness, or accepts an explicit target, and writes each skill as a direct child of that root. It reports the root it wrote and exits nonzero when a target is incomplete.

Install for all detected harnesses

npx --yes --package @youcisla/agent-foundry agent-foundry-install

List targets

node scripts/install.js --list

Dry run one target

node scripts/install.js --harness=codex --dry-run

Run the installer from the repository clone. The npm package exposes the same installer as the agent-foundry-install binary.

Install the runtime (optional)

The runtime is a Python package (Python 3.10 to 3.14). It provides the CLI, the local daemon, the SQLite execution log, and the MCP stdio server. The distribution contains no skill content; install the corpus with the installer above.

Install the runtime

pip install agent-foundry-runtime

Run the MCP server without installing

uvx agent-foundry-runtime mcp

Provider configuration is local. Set the key your model configuration expects in the environment or config file; agent-foundry doctor reports whether a key is present without printing it.

Verify discovery

For skills-only use, confirm that the harness root contains one directory per skill:

Example for Claude Code

ls ~/.claude/skills | head

For the runtime, the doctor command checks configuration, the index, the daemon, and the provider key. A failure exits nonzero so it can gate a setup script.

Health check

agent-foundry doctor

Rebuild the skill index

agent-foundry index

MCP server

The MCP server speaks JSON-RPC over stdio and exposes the same catalog operations as the CLI: list and search skills, get one skill, plan, execute, verify, classify an action, read runtime status, and read execution history.

Example client configuration

{
  "mcpServers": {
    "agent-foundry": {
      "command": "uvx",
      "args": ["agent-foundry-runtime", "mcp"]
    }
  }
}

MCP clients start the server themselves; no daemon port has to be open.

Local daemon

agent-foundry serve starts a FastAPI daemon on 127.0.0.1:8765. Most CLI commands start it lazily when needed. Requests need a bearer token stored in a user-owned file, and the daemon refuses non-loopback binding unless explicitly configured.

The CLI currently exposes 25 commands.

Start the daemon

agent-foundry serve

Check liveness

curl -s http://127.0.0.1:8765/health

Configuration and retention

Configuration lives in ~/.config/agent-foundry/config.toml (platform-aware). Executions are stored in a local SQLite database. Content retention can be disabled; pruning is explicit and supports a dry run.

Write the default config

agent-foundry init

Inspect retention first

agent-foundry prune --dry-run

Security boundary

  • The daemon binds loopback, requires a bearer token for requests, and rejects unexpected Host and Origin headers.
  • Provider keys are read from your environment or config and are never printed by doctor.
  • Policy checks can classify a shell, file, text, or network action as allow, confirm, or deny before it runs.
  • The public site has no telemetry and no third-party runtime JavaScript.

Troubleshooting

SymptomCheck
No configuration foundRun agent-foundry init, then agent-foundry doctor.
Index is emptyInstall the skill corpus, then run agent-foundry index.
Daemon unreachableRun agent-foundry serve in another terminal; check the printed port and token path.
Provider errorConfirm the key and model id in your local configuration; doctor reports only presence, not the value.
MCP client does not startConfirm uvx agent-foundry-runtime mcp runs in your shell, then copy the JSON block above.

Development documentation and the full source live in the repository. Install notes for every harness are in INSTALL.md and docs/harness-compatibility.md.