@render/pi-render

Render integration for the Pi coding agent — Render's OSS skills plus hosted Render MCP access, in one installable package.

v1 is intentionally a subset of the sibling render-plugin-claude-code, render-opencode-plugin, render-codex-plugin, and render-cursor-plugin family: it ships the skills and the control plane, not the slash commands, subagent, or Blueprint hook.

Install

pi install https://github.com/render-oss/render-pi-plugin

Pi installs the package's runtime dependencies, including its MCP adapter. There is no separate MCP package or MCP support prerequisite to install.

For a reproducible install pinned to this release:

pi install https://github.com/render-oss/render-pi-plugin@v0.1.0

What you get (v1)

• Skills synced from render-oss/skills — Render deployment, Blueprint, debugging, and monitoring know-how, auto-loaded by the model and invocable as /skill:render-*.
• Render MCP access via pi-mcp-adapter (OAuth by default, RENDER_API_KEY for CI) — the full control plane: services, deploys, logs, env vars, Postgres.

MCP is lazy and proxy-only. On a clean first run, the adapter makes one best-effort Render connection to populate its local tool-metadata cache, but it does not open an OAuth flow. If that bootstrap cannot connect, Pi and the Render skills continue to work; authenticated MCP operations connect on demand.

Because that bootstrap runs before you have credentials, the first session reports once that the Render server requires authentication (an HTTP 401). This is expected, not a failure — Pi and all 21 skills load normally. Run /mcp-auth render when you want the Render tools, and the notice stops.

Auth

• OAuth (recommended, interactive): run /mcp-auth render in a session — no API key needed.
• RENDER_API_KEY (CI/non-interactive): set the env var before starting Pi; the adapter reads it only when making a lazy connection and sends it as a bearer token.

The Render server is configured by this package rather than read from your MCP config, so it stays isolated from any MCP servers you've set up yourself. The trade-off: /mcp setup, /mcp enable, and /mcp disable don't apply to it, and /mcp status reports what this package configured.

Headless Linux OAuth requires an unlocked system credential store. Use RENDER_API_KEY when one isn't available.

Skills-only mode

Run pi config and disable this package's extension while leaving its skills enabled. This removes the MCP tools but keeps every /skill:render-* command available.

Update or remove

pi update --extensions
pi remove https://github.com/render-oss/render-pi-plugin

An install pinned to @v0.1.0 stays pinned; install a newer tag explicitly to move it forward.

Compatibility

v0.1.0 is verified against Pi 0.83.0, pi-mcp-adapter 2.19.0, and Node.js 22.19.0 or newer. See docs/SPEC.md for the complete scope and deferred work.

Development

npm install
npm run verify   # typecheck + lint + tests — the single green light

Test-first; see docs/SPEC.md §7.

Updating skills

renderinc/skills is the source of truth. Copybara mirrors its distributable files to render-oss/skills, and this repository syncs from that public mirror. .github/workflows/sync-skills.yml checks daily at 06:00 UTC and can also be run manually; it verifies changes before opening a skills-sync pull request.

Run the same sync locally:

npm run sync-skills
npm run sync-skills -- --ref <commit-or-tag>
npm run sync-skills -- --repo ../../skills --ref main  # local source-of-truth clone
npm run verify

The script applies the public mirror's source-only eval exclusions and records the resolved source, ref, and commit in skills/.sync-source. Never edit vendored skills by hand. Keep auto-merge disabled unless the repository requires the verify check; then set the SKILLS_SYNC_AUTOMERGE=true repository Actions variable to allow squash auto-merge.

The public mirror is the release input. Public contributions can temporarily lead the internal clone until their reverse-Copybara PR is merged, so compare local-clone output with render-oss/skills before using it in a release.