GitHub
Connect AI assistants to GitHub - manage repos, issues, PRs, and workflows through natural language.
作者提供的说明(英文原文)
Bounded, source-write-free Rust correctness tools grounded in Cargo and rustc output.
这三项永远不会合并成一个分数:意见、测量和人气是不同的东西。
我们还没有为它写通俗的介绍。下面是我们从注册中心读取到的事实。
选择你使用的 AI 应用,然后复制设置。
作者没有在注册中心列出安装信息。请查看下方作者的 README。
它在你的电脑上运行;安装后,工具列表会显示在你的 AI 应用中。
我们检查的是任何人都能看到的信息;我们没有阅读代码。不适用的检查项不计入总数。 · 检查于 2026年11月8日
注册名称已通过 GitHub 账户 ugur-murat-alt 验证,该账户拥有此仓库
我们还没有从 GitHub 读取这个仓库
我们还没有从 GitHub 读取这个仓库
它在你的电脑上运行,没有互联网端点
在你的电脑上运行,改用本地设置
它在你的电脑上运行;我们只能读取互联网端点的工具列表
与注册中心里这个分类中所有活跃的服务器比较。只有当发布者拥有该仓库时,星标数才计入。
它在 1 项适用的基础检查中通过了 1 项。我们没有阅读代码。在添加密钥或令牌之前,请先看看这个服务器要求哪些权限,并使用只具备所需权限的密钥。
注册中心没有列出必需的密钥,不过有些服务器会在首次使用时要求你登录。
它在你的电脑上运行。你的 AI 应用会在需要时启动它,数据不会经过作者的服务器(除非它自己调用在线服务)。
GitHub 仓库最后更新于–,最新版本(0.2.0)于2个月前进入注册中心。
支持 MCP 的应用,例如 Claude Desktop、Claude Code、Cursor 和 VS Code。你的电脑上还需要有 Rust。
同一分类中有:GitHub, Desktop Commander, Cloudflare MCP
同一分类中其他作者的服务器,按星标数排序。
如何看卡片:★ GitHub 星标i三大支柱:评价 · 检查 · 使用情况i更新 = 最新版本发布到注册中心的时间i
Connect AI assistants to GitHub - manage repos, issues, PRs, and workflows through natural language.
MCP server for terminal commands, file operations, and process management
Cloudflare MCP servers
按原文从作者的仓库复制。PopNx MCP 没有审查或背书其内容。HTML、图片和脚本已被移除。链接会离开本站。
English | Türkçe
An AGZ Yazılım product. agz-rust-mcp is a standalone MCP server with stdio and Streamable HTTP transports for general-purpose Rust development grounded in Cargo and rustc. Discover a project offline, collect focused source context and compiler diagnostics, select individual tests, and run binaries/examples. Exact-version dependency docs, Rust Analyzer navigation and write-free change/refactor packages support deeper work.
The same workflow works with a library, CLI, service or multi-package Cargo workspace. No framework-specific project structure or particular MCP client is required.
| Need | Tool call |
|---|---|
| Map packages, declared targets, features and dependencies without compiling | workspace(action="inspect") |
| Locate code and read bounded, numbered source | workspace(action="search", query="parse_record"), then workspace(action="read", path="src/lib.rs", line=1, lineCount=80) |
| Get fresh compiler diagnostics | check(target="check", options={"packages":["my_crate"]}) |
| Discover actual test names | test(action="list", packages=["my_crate"]) |
| Run one library test by its full name | test(action="run", packages=["my_crate"], cargoTarget={"kind":"lib"}, filter="tests::parses_record", exact=true) |
| Compile tests without executing them | test(action="compile", packages=["my_crate"]) |
| Execute a binary or example | run(package="my_crate", target={"kind":"example","name":"demo"}, args=["input.json"], timeoutMs=30000) |
| Resolve an unfamiliar dependency API | docs with its exact version; api for deeper resolution/probing |
Pass an absolute dir for the active project when the client starts elsewhere. A function is exercised through a test or executable example. run is a typed Cargo runner; program arguments remain literal, including spaces and shell metacharacters. Tests and applications execute project code with its ordinary side effects. Empty selections return NO_TESTS, not a passing test verdict.
Use agz-rust-mcp --toolset core for a compact catalog of six everyday tools: workspace, check, test, run, docs, and crate_lookup. The default --toolset full preserves every existing tool plus the new additions. Individual --tools- false switches can disable tools further. See the full reference for limits and result states.
# installer script (Linux x86_64): download install.sh and SHA256SUMS from the
bash install.sh
cargo install agz-rust-mcp --locked agz-rust-mcp --version
Prebuilt .tar.gz archives with .sha256 files are available for Linux x86_64, macOS arm64, and Windows x86_64. Releases up to 0.2.0 use the legacy agz-rust-coder-* asset names; from 0.3.0 onward the names are agz-rust-mcp-*.
Step-by-step instructions, per-OS notes, checksum verification, and troubleshooting: [docs/install.md](docs/install.md) · [Türkçe](docs/install.tr.md).
Use the installed executable with any client supporting local stdio MCP. Supply these fields in the client's own server configuration format:
{
"command": "agz-rust-mcp", "args": ["--allow-root", "/absolute/path/to/project"], "env": {"AGZ_RUST_MCP_SERVER__TOOLSET": "core"} }
For a client that uses HTTP, start the same binary as a local Streamable HTTP server and configure http://127.0.0.1:8787/mcp as its MCP URL:
agz-rust-mcp --transport http --http-bind 127.0.0.1:8787 \
--allow-root /absolute/path/to/project
The HTTP listener is loopback-only. It supports the session lifecycle used by 2025-11-25 clients and the stateless 2026-07-28 protocol. Client-side display prefixes, configuration file locations, and automatic prompt selection belong to the client; the server depends on none of them. The canonical working directory is the default authorized root; client-provided roots can only narrow configured access. See connection details.
Four offline skills ship inside the executable: agz-rust-workflow, agz-rust-repair, agz-rust-refactor, and agz-rust-performance. They are available immediately through all three standard MCP interfaces:
workspace(action="skills") lists the catalog; workspace(action="skills", skill="agz-rust-workflow") reads a skill.prompts/list and prompts/get expose workflow, repair, refactor, and performance.resources/read exposes agz-rust-mcp://skills/ .A client that only exposes tools can use the complete skill content. Version and SHA-256 metadata identify the exact bundled instructions. No skill installation or separate MCP is required. Optional agz-rust-mcp skills export --dir creates files for clients that also support filesystem skills; existing files are never overwritten. See built-in skills.
Read the documentation in this order:
The bilingual index with the full reading path is docs/README.md / docs/README.tr.md.
| Contract | Value |
|---|---|
| Crate, binary, server | agz-rust-mcp |
| MCP Registry | io.github.agzyazilim/rust-mcp |
| Source version | 0.6.0 |
| First release | 0.1.0 |
| Release tag | agz-rust-mcp-v |
| Rust edition / MSRV | 2024 / 1.91.0 |
| Rust MCP SDK | rmcp 3.8.0 |
| Default / discovered protocol | 2025-11-25 / 2026-07-28 |
MCP package ownership marker: mcp-name: io.github.agzyazilim/rust-mcp.
These are the exact server tool names. Use the names returned by tools/list; a client may display its own namespace or prefix.
| MCP tool | Default | Purpose |
|---|---|---|
workspace | enabled | Read built-in skills, inspect manifests/lockfiles, search source and read excerpts offline. |
test | enabled | Discover, compile or run selected tests with exact names, ignored-test controls and bounded execution. |
run | enabled | Run an exact Cargo binary/example with literal arguments, a deadline and bounded output. |
check | enabled | Run bounded Cargo check, Clippy, tests, docs, or the full gate. |
profile | enabled | Analyze observed Cargo rebuild behavior and compare bounded build evidence without claiming unmeasured speedups. |
audit | enabled | Scan Rust source for bounded static findings. |
crate_lookup | enabled | Verify a crate and optional exact version on crates.io. |
docs | enabled | Resolve exact-version docs from cache, local sources, or docs.rs. |
context | enabled | Prepare, expand, or delta a revision-bound semantic context capsule with per-item reasons. |
api | enabled | Resolve an API signature from bounded analyzer evidence or type-check a candidate snippet in an isolated copy of the workspace configuration. |
explain | enabled | Explain macro expansion provenance, trait obligations, or cfg enablement. |
verify | enabled | Plan or run a bounded feature, target, toolchain, and stage matrix. |
symbol | enabled | Read Rust Analyzer hover data for one symbol. |
references | enabled | Find bounded references. |
definition | enabled | Find the selected definition. |
symbols | enabled | List symbols in one Rust file. |
implementations | enabled | Find implementations. |
hierarchy | enabled | Trace a bounded call hierarchy. |
rename | enabled | Produce a verified rename edit package without applying it. |
refactor | enabled | Produce a verified refactor edit package without applying it. |
change | enabled | Create, stage, migrate, and validate a revision-bound changeset in server-owned scratch without writing the workspace. |
repair | enabled | Analyze, try, compare, and minimize compiler-driven repair candidates for a failing change revision without writing the workspace. |
work | enabled | Drive a typed intent through change/validate with explicit gates and budgets, returning honest gate evidence or a bounded single-use handoff. |
Every tool returns deterministic structured data plus an equivalent bounded text fallback. External data stays under untrustedData. Expected domain outcomes such as compiler failure, missing crates, or unavailable docs are typed results; invalid input, authorization failure, resource exhaustion, and unavailable semantic infrastructure are protocol errors.
Precedence is CLI, then AGZ_RUST_MCP_* environment variables, then the explicit --config TOML file, then defaults. Environment keys use __ between sections, for example AGZ_RUST_MCP_GATE__HARD_TIMEOUT_MS=600000.
| Key | Default | Meaning |
|---|---|---|
server.transport | stdio | MCP transport: stdio or http. |
server.http_bind | 127.0.0.1:8787 | Loopback address for the HTTP /mcp endpoint. |
server.allow_roots | canonical CWD | Workspace read/command boundary. |
server.allow_dependency_roots | empty | Explicit external path-dependency roots. |
gate.hard_timeout_ms | 600000 | Cargo operation deadline. |
gate.scope | shadow | Validation target: workspace, shadow, or affected. |
gate.cache | auto | Cache policy: auto, project, or isolated. |
rust_analyzer.workspace_code | deny | Reject RA startup unless workspace code is disabled. |
docs.fallback | auto | Documentation source policy. |
profile.max_report_bytes | 4194304 | Bounded read/store cap for one Cargo timing artifact. |
profile.max_runs | 4 | Fresh Cargo runs available to one profile call. |
profile.compare_samples | 3 | Required samples per side before any speed claim. |
limits.tool_output_bytes | 49152 | Maximum serialized tool result size. |
change.max_bytes | 268435456 | Maximum captured candidate bytes per changeset. |
repair.max_candidates | 4 | Candidates one repair action may try. |
repair.max_compiles | 4 | Cargo validations one repair action may run. |
repair.wall_time_ms | 120000 | Wall-clock budget for one repair action. |
repair.minimize_max_candidates | 32 | Compile-evaluated reduction attempts one repair(action=minimize) may try. |
repair.minimize_max_compiles | 16 | Cargo runs one repair(action=minimize) may execute, including reproduction and export verification. |
work.max_candidates | 4 | Host candidate revisions one work item may stage. |
work.max_compiles | 12 | Gate validations one work item may run. |
work.wall_time_ms | 600000 | Wall-time budget for one work item. |
telemetry.enabled | true | Bounded local activity records without prompts or source. |
profile evidence is bounded: the latest 64 records stay in memory, and persisted timing artifacts under the server-owned profile-evidence directory are pruned after 24 hours or beyond 64 files. The profile result reports this retention policy, so exp
只显示了开头部分;其余内容请在仓库中阅读