MCP SERVER · RUST TOOLCHAIN
nxm-mcp-exec
An MCP server for typed Rust toolchain diagnostics — fewer tokens, better code.
Built to run entirely locally. Open source under the MIT license.
Why I built it
I built nxm-mcp-exec to give a coding agent a better grip on a Rust codebase: higher code quality, and far fewer tokens burned per check.
A normal agent runs cargo in a shell and reads back hundreds of lines of human-formatted log text — then spends tokens (and makes parsing mistakes) turning that text into something it can act on. nxm-mcp-exec removes that cost. It runs the command for the agent and returns a small, typed result — { ok, diagnostics, summary } — instead of a wall of logs. The agent reads a compact, structured answer, spends a fraction of the tokens, and never writes a log-parsing bug again.
The result: cleaner Rust code, a tighter feedback loop, and a much smaller token footprint inside the agent’s context window.
What it is
nxm-mcp-exec is a small, security-first MCP server. It runs allow-listed Rust toolchain commands inside a target repository and returns structured, typed diagnostics over stdio JSON-RPC. It is not a CI — it is a diagnostics surface built for coding agents. Every call is authenticated and every repo path is confined before a single cargo command is spawned.
Typed, not log text
Every tool returns { ok, …, summary } — structured JSON an agent can act on directly. No screen-scraping, no log-parsing bugs, a fraction of the tokens.
Security-first by construction
Two gates run before any command: authentication (nxm_user / nxm_token) and repository-path confinement. Only an enumerated allow-list of cargo subcommands can ever run — no arbitrary shell.
One GO / NO-GO verdict
run_gate composes the individual checks into a single, short-circuiting verdict. The strict gate adds security and dependency-hygiene checks on top of the default correctness sweep.
How it works
Architecture — three layers, two security gates before any cargo command runs
Tool catalogue — nine typed tools over one strict allow-list
The strict gate — six ordered checks, short-circuit at the first failure
The nine tools
cargo_check
Typed compiler diagnostics
cargo_build
Full build (release, features)
cargo_test
Test results (nextest or libtest)
cargo_clippy
Lint diagnostics, optional deny_warnings
cargo_fmt_check
Formatting status (read-only)
cargo_audit
Security advisories — fail-closed
cargo_machete
Unused dependencies
cargo_outdated
Out-of-date deps — informative only
run_gate
A single GO / NO-GO verdict composing the checks
cargo_outdated is the only informative tool — it reports, it never fails the gate.
Created by Daniele Oddo — AI Engineering Architect
Project: nxm-mcp-exec · github.com/dangranaz/nxm-mcp-exec
View on GitHub →
Tech details
Language
Rust (edition 2021)
Transport
JSON-RPC 2.0 over stdio
serde / serde_json
parses cargo --message-format=json
anyhow
error handling — no unwrap/expect in production
tracing
structured logs on stderr
nxm-mcp-authz
shared auth client — never re-implemented
Deploy
single static binary, no runtime
No unwrap/expect in production. Allow-listed cargo subcommands only — no arbitrary shell. Two gates before any command: authentication (nxm_user / nxm_token) and repository-path confinement.