Architecture

A small static site, a Markdown corpus, and an optional local runtime. Nothing on this site executes third-party JavaScript, and the runtime never calls home.

How it fits together

Architecture: packages feed the installer, which writes skills and agents into harness roots; the local runtime exposes a daemon and MCP server and logs to SQLitenpm / PyPI / MCPInstallerSkills and agentsHarness rootsLocal runtimeFastAPI daemonMCP stdioSQLite log
Packages and the installer write skills into harness roots; the runtime is independent.
Text version of the diagram
  1. The npm, PyPI, and MCP registry channels publish the same version.
  2. The installer writes skills and agents as direct children of a harness root.
  3. Each harness discovers the Markdown skills with its own rules.
  4. The optional Python runtime serves a local daemon and an MCP stdio server.
  5. Execution metadata is stored in a local SQLite database.

The site itself is generated from the same records: a Python build script renders static HTML, CSS, one small JavaScript file, inline SVG diagrams, a catalog manifest, and a sitemap.

Channels

npm

@youcisla/agent-foundry installs the skill corpus and the harness installer. The runtime has its own Python distribution.

PyPI

agent-foundry-runtime provides the CLI, the local daemon, and the MCP stdio server. It contains no skill content; install the corpus separately.

MCP registry

io.github.youcisla/agent-foundry points at the same runtime server so MCP clients can start it without a manual command.

Components

Corpus

Skills under skills/core and skills/optional, each <id>/SKILL.md with YAML frontmatter. Agents under agents/<id>/AGENT.md. The installer requires each skill to be a direct child of the harness root.

Runtime

A Python package with a FastAPI daemon bound to loopback, a SQLite execution log, a deterministic indexer, and an MCP server. The daemon is started lazily by CLI commands.

Site build

One generator reads the tracked corpus and a metrics file, renders safe Markdown and static pages, and publishes a manifest of hashed output files. Runtime JavaScript on the site is limited to navigation, filtering, and copying commands.

Trust boundary

Execution content, provider keys, and local paths stay outside the site. The published directory is allowlisted and checked for personal paths and receipt-like files.

Harness roots

AdapterRoot
Claude Code~/.claude/skills/<skill>
Codex CLI~/.agents/skills/<skill>
Cursor~/.cursor/skills/<skill>
Hermes~/.hermes/skills/agent-foundry/<skill>
Gemini CLI~/.gemini/skills/<skill>
OpenCode~/.config/opencode/skills/<skill>
Antigravity~/.gemini/config/skills/<skill>
CodeBuddy~/.codebuddy/skills/<skill>
JoyCode~/.joycode/skills/<skill>
Qwen Code~/.qwen/skills/<skill>
Kimi CLI~/.kimi-code/skills/<skill>
OpenClaw~/.openclaw/skills/agent-foundry/<skill>
Claude Code (project)<project>/.claude/skills/<skill>
Zed~/.agents/skills/<skill>

Roots are maintained by the installer; run node scripts/install.js --list for the current list.

Version 0.6.0. Read the documentation for setup and configuration, or the source for implementation details.