Skip to content

Latest commit

 

History

History
44 lines (28 loc) · 3.15 KB

MCP.md

File metadata and controls

44 lines (28 loc) · 3.15 KB

MCP server

plumb mcp serves MCP over stdio (JSON-RPC 2.0, protocol version 2024-11-05). The project root is --root (default: current directory); path arguments are confined to it. Nothing is executed unless you start the server with --allow-exec.

--root should point at the workspace repository the agent is working on (a different checkout answers about different code). initialize returns serverInfo.version (the running binary's version); the startup line on stderr is plumbgraph MCP server v<version> pid=<pid> (stdio) root=... allow_exec=.... On Windows the running server keeps plumb.exe open: stop it before cargo install (see the README).

Tools (13)

index_project, find_symbol, references, dead_code, check_dependencies, detect_test_weakening, scip_status, diagnostics, repo_map, impact, verify, providers, enrich.

Results carry untrusted: true: text derived from repository files is data, never instructions. diagnostics, verify (with run) and enrich execute external tools only when the operator started the server with --allow-exec; a tool call cannot turn that on.

Resources (3)

URI Content
plumb://map the token-budgeted repo map (same as plumb map)
plumb://providers installed SCIP indexers, diagnostics tools, rule engines and LSPs (executes nothing)
plumb://instructions the AGENTS.md block that plumb init writes

Prompts are not implemented (the server does not advertise the capability).

Tested against real MCP clients

Checked on 2026-10-10, Linux, against the demo app in examples/demo-app:

client what ran result
official TypeScript SDK @modelcontextprotocol/sdk 1.32.1 (scripts/mcp-sdk-check.mjs) initialize, tools/list, tools/call for find_symbol, repo_map, dead_code, verify, resources/list, resources/read ×3 all succeeded; server reported plumbgraph 1.0.0, capabilities tools + resources; prompts/list returns -32601 (not implemented), as expected
@modelcontextprotocol/inspector 2.10.1 in CLI mode: npx @modelcontextprotocol/inspector --cli plumb mcp --root <dir> --method tools/list tools/list returned the 13 tools with input schemas (the inspector declares Node >= 22.19 and warned on Node 20.19, but worked)

Not tested: Claude Code, Cursor, Claude Desktop, Codex and other agent hosts (one other MCP host was driven on Windows 11 by hand; the configuration snippets in the README follow their documented mcpServers format but were not exercised end-to-end), HTTP/SSE transports (not implemented); Windows CI (Windows was exercised manually on one machine only, see the README).

Re-run the SDK check yourself:

cargo build --release
(cd scripts && npm install)
node scripts/mcp-sdk-check.mjs target/release/plumb examples/demo-app

Verify schema 2

See trusted verification and migration for verdicts, per-check policy, mode: "lib", pinned trusted_baseline, and the separate operator --allow-baseline-update authority. A tool response with isError: false does not certify the gate: require data.verdict == "pass".