project-mcp-json

A pi extension that reads a Claude Code-style .mcp.json from the project working directory and registers the configured MCP servers with pi via pi.registerMcpServer.

Requires Pi 1.0 or later. The pi.registerMcpServer API used by this extension was introduced in Pi 1.0. On older Pi versions the extension loads silently and emits a one-time notice instead of registering any servers.

Why

Different AI coding environments use different files to configure MCP servers:

Environment	Config file
Pi	~/.pi/agent/mcp.json (user) and .pi/mcp.json (project, trusted)
Claude Code	.mcp.json (project root)
Codex	~/.codex/config.toml (user) and .codex/config.toml (project, trusted)

If your project already has a .mcp.json for Claude Code, this extension lets Pi use it directly — no duplication required.

Pi's own mcp.json takes precedence by name. If a server name defined in Pi's own mcp.json (~/.pi/agent/mcp.json or .pi/mcp.json) matches one this extension registers from .mcp.json, Pi's own entry wins — the existing server is not replaced and no collision error is raised. This extension is most useful when Pi has no mcp.json of its own, or when the server names do not overlap.

Discovery

The extension reads only <cwd>/.mcp.json — it does not walk parent directories. If the file is absent the extension is a no-op for that session.

Trust model

The extension follows Pi's project trust semantics for deciding whether to load .mcp.json. It treats .mcp.json as a trust-requiring project resource, mirroring the logic Pi uses for its own .pi/ config resources.

Decision order (evaluated each session)

1. Project not trusted (/trust has not been granted) → skip silently.
2. Project has Pi resources (e.g. .pi/mcp.json, .pi/extensions/, .agents/skills/ in the project or an ancestor) → Pi already made a genuine trust decision for this project; load .mcp.json as well.
3. Saved trust decision — check ~/.pi/agent/trust.json for the nearest ancestor of the project directory:
   • true → load.
   • false → skip silently.
4. defaultProjectTrust global setting (from ~/.pi/agent/settings.json):
   • "always" → load.
   • "never" → skip silently.
   • "ask" (default) → skip; if interactive UI is available, emit one info notification: "Found .mcp.json — run /trust to enable its MCP servers."

Enabling .mcp.json loading

The standard way is to run /trust inside a Pi session in the project directory. This saves true for the project path in trust.json (step 3 above) and causes future sessions to load .mcp.json automatically.

Alternatively, set defaultProjectTrust to "always" in Pi's global settings (step 4) to enable .mcp.json loading for all trusted projects.

Note on --approve / one-shot trust flags: Pi's --approve flag is not detectable from within an extension. It grants a one-shot project-trust session but does not write to trust.json, so this extension behaves as if no saved decision exists and falls through to step 4. Run /trust inside the session to persist the decision.

Parent-directory trust and nearest-ancestor semantics

Saved trust decisions in trust.json use nearest-ancestor matching: the extension walks up the directory tree from cwd and uses the first decision it finds. If you trust a parent directory (via /trust → "Trust parent folder"), all projects under it — including their .mcp.json files — are loaded by this extension.

Undo / revoke

Reliable persistent denial means saving an explicit false for the project path. The most direct way is to run /trust inside the project and choose "Do not trust" — Pi writes false for the exact project path to ~/.pi/agent/trust.json and the extension will skip .mcp.json even if a parent directory is trusted.

Other approaches and their limitations:

• Delete the entry (or set it to null) in trust.json: this resumes nearest-ancestor lookup (step 3), not a direct fallback to defaultProjectTrust. If any ancestor directory has a saved true, that decision is found first and loading is re-enabled. defaultProjectTrust (step 4) only applies when the nearest-ancestor walk finds no decision at all. Deleting/nulling an entry is therefore not a reliable denial when any ancestor path is trusted.
• Set defaultProjectTrust to "never" in ~/.pi/agent/settings.json: this prevents loading only when the nearest-ancestor walk finds no saved decision (neither true nor false). A saved true for the project or any ancestor takes precedence; use an explicit saved false for the project path if both coexist.
• Manually edit trust.json: ~/.pi/agent/trust.json is a plain JSON object mapping absolute paths to true/false/null. Set the project path to false to permanently deny; set it to null or delete the key to resume nearest-ancestor lookup (a trusted parent may then re-enable loading). For reliable denial, always save false explicitly.

Supported subset

The extension maps a subset of Claude Code's .mcp.json schema to Pi's McpServerConfig. Fields and features that are accepted:

• stdio servers: command, args, env, cwd, timeout, description
• http / streamable-http servers: url, headers, oauth (clientId, clientSecret, callbackPort, callbackUrl, scope), timeout, description
• ${VAR} and ${VAR:-default} environment variable expansion in command, args, env values, cwd, url, headers values, and oauth.clientSecret

Features and fields that are not imported:

Skipped feature	Reason
sse server type	Not part of Pi's McpServerConfig
ws / websocket server type	Not part of Pi's McpServerConfig
headersHelper field	Rejected by strict field allowlist
auth field	Rejected by strict field allowlist
Claude Code approvals (.claude/settings.json)	Not read; Pi manages its own trust
Claude Code OAuth tokens	Not imported; Pi manages its own OAuth state

Unknown top-level fields cause the individual server entry to be skipped with a warning (not the whole file). sse and ws type strings and auth/headersHelper fields always produce a skip warning.

Secrets caveat

Environment variable values in env and headers are expanded at session start and then encoded as literals before being passed to pi.registerMcpServer. After registration, expanded values may be visible in Pi's /mcp panel or wherever Pi surfaces server configuration. To reduce secret exposure:

• Prefer environment variables over literal values in .mcp.json. The extension expands ${VAR} at startup but the raw variable name is what appears in .mcp.json (which is often checked into source control), not the value.
• Use headers for auth tokens on HTTP servers rather than embedding secrets in the URL.
• Add .mcp.json to .gitignore if it contains any literal secrets.

Install

Standalone npm package

pi install npm:@diegopetrucci/pi-project-mcp-json

Collection package

pi install npm:@diegopetrucci/pi-extensions

GitHub package

pi install git:github.com/diegopetrucci/pi-extensions

Then reload pi:

/reload

Example .mcp.json

{
  "mcpServers": {
    "my-local-tool": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@my-org/my-mcp-server"],
      "env": {
        "API_KEY": "${MY_API_KEY}"
      }
    },
    "my-remote-tool": {
      "type": "http",
      "url": "${TOOL_SERVER_URL}",
      "headers": {
        "Authorization": "Bearer ${TOOL_TOKEN}"
      }
    }
  }
}