workbench-mcp
workbench-mcp is a Python FastMCP server for controlled local workspace inspection, bounded file reads/searches, guarded text patching, allowlisted test execution, read-only Git status, approved artifact collection, and workspace diagnostics.
README
workbench-mcp
workbench-mcp is a Python FastMCP server for controlled local workspace inspection,
bounded file reads/searches, guarded text patching, allowlisted test execution, read-only
Git status, approved artifact collection, and workspace diagnostics.
It is intended for developer workbenches where an MCP client needs useful repository context without receiving unrestricted filesystem or shell access.
Problem Statement
AI-assisted development tools often need to inspect files, run tests, and collect reports. Giving those tools raw shell access or broad host filesystem access is risky. This project wraps common development actions in typed MCP tools backed by explicit workspace, artifact, command, timeout, output, and redaction controls.
Project Status
Status: v0.1.0 release preparation on branch codex/workbench-mcp-v1.
Implemented and locally verified:
- stdio FastMCP server startup.
- In-memory and stdio MCP client/server smoke tests.
- Docker image build, direct container smoke test, Compose config, and Compose smoke workflow.
- Reproducible public demo that creates a temporary Git workspace and writes a JSON report.
- GitHub Actions CI is green for the latest pushed Phase 7 commit
34fc81ae32e4404f01f3f460b43e6508e108421din run29778875177. - Local final-audit fixes after that pushed commit require a new CI run after push.
Not implemented:
- HTTP or Streamable HTTP transport for this project.
- Complete arbitrary-code execution sandboxing.
- Package publishing, Docker image publishing, Git tags, or GitHub releases.
Feature Overview
- Workspace metadata through
workspace_info. - Directory listing and UTF-8 text reads bounded to
WORKSPACE_ROOT. - Text search with glob, case-sensitivity, result, and output limits.
- One guarded expected-content text replacement through
apply_patchwhenREAD_ONLY_MODE=false. - Allowlisted command execution using argument arrays and
shell=False. - Named predefined test commands with JSON reports under the artifact root.
- Read-only Git status using non-mutating Git commands.
- Approved artifact collection from category-specific artifact directories.
- Diagnostic findings for workspace, artifact, executable, Git, and transport state.
- Secret-pattern and host-root redaction before MCP-facing responses.
Architecture
flowchart TD
Client["MCP client"] --> Transport["FastMCP stdio transport"]
Transport --> Server["workbench_mcp.server.create_server"]
Server --> ToolLayer["Typed MCP tool/resource layer"]
ToolLayer --> Services["Service layer"]
Services --> Config["Validated WorkbenchConfig"]
Services --> Security["Security helpers"]
Security --> Workspace["WORKSPACE_ROOT boundary"]
Security --> Artifacts["ARTIFACT_DIRECTORY boundary"]
Services --> Process["subprocess with shell=False"]
Services --> Git["read-only git status"]
ToolLayer --> Errors["safe ToolError conversion and redaction"]
More detail: docs/architecture.md.
Registered MCP Tools
These are the exact tool names registered by src/workbench_mcp/tools/common.py and
verified by tests/unit/test_mcp_server.py.
| Tool | Inputs | Behavior |
|---|---|---|
workspace_info |
none | Returns sanitized workspace metadata, read-only status, limits, capabilities, and Git summary. |
list_files |
relative_directory=".", max_depth=2, pattern=null, max_results=1000 |
Lists workspace-contained files/directories/symlinks with depth and result limits. |
read_file |
relative_path, optional start_line, optional line_count |
Reads a workspace-contained UTF-8 text file within size limits. |
search_text |
query, relative_directory=".", glob=null, case_sensitive=false, result_limit=100, max_output_bytes=null |
Searches workspace text files with output and result bounds. |
apply_patch |
relative_path, expected_content, replacement_content |
Replaces exactly one expected text block when writes are enabled. |
run_command |
command, cwd="." |
Runs an allowlisted executable with argument-array execution. |
run_tests |
test_name |
Runs a configured named test command and writes a test report. |
git_status |
none | Returns branch, staged/modified/untracked files, and diff statistics without mutating Git state. |
collect_artifact |
category, relative_path |
Reads approved text artifacts from test_report, coverage_report, structured_log, or diagnostic_report. |
diagnose_workspace |
none | Returns structured findings with severity, evidence, probable cause, and remediation. |
No destructive Git tools, delete-file tools, shell tools, or unrestricted command tools are registered.
Server-Information Resource
Exact resource registration:
| Field | Value |
|---|---|
| Name | server_info |
| URI | workbench://server-info |
| MIME type | application/json |
Payload keys returned by server_information_payload:
{
"package_version": "0.1.0",
"server_capabilities": {
"workspace_inspection": true,
"filesystem_read": true,
"text_search": true,
"controlled_text_patch": false,
"allowlisted_commands": [],
"predefined_tests": [],
"git_status": true,
"artifact_collection": true,
"diagnostics": true
},
"registered_tool_names": [
"workspace_info",
"list_files",
"read_file",
"search_text",
"apply_patch",
"run_command",
"run_tests",
"git_status",
"collect_artifact",
"diagnose_workspace"
],
"active_safety_limits": {
"max_file_size_bytes": 1048576,
"max_command_seconds": 30,
"max_output_bytes": 1048576
},
"read_only_status": true,
"sanitized_workspace_metadata": {
"name": "workbench-mcp",
"exists": true,
"is_directory": true
},
"supported_transports": ["stdio"]
}
Values reflect the active configuration. The tool list and supported transports are fixed for this release.
Installation With uv
Install uv, then install the locked project environment:
uv sync --frozen --all-groups
uv run python --version
uv run workbench-mcp --version
On this Windows verification shell, bare uv is not on PATH; the equivalent verified
command prefix is:
& $env:APPDATA\Python\Python313\Scripts\uv.exe run python --version
Configuration Reference
Configuration loads from defaults, optional TOML via WORKBENCH_MCP_CONFIG, and
environment variables. Environment variables override TOML values.
| Environment variable | TOML key | Default | Notes |
|---|---|---|---|
WORKBENCH_MCP_CONFIG |
N/A | unset | Optional TOML file path. The file may contain a [workbench_mcp] table or top-level keys. |
WORKSPACE_ROOT |
workspace_root |
current working directory | Must exist and be a directory. All workspace paths must resolve inside it. |
READ_ONLY_MODE |
read_only_mode |
true |
Blocks apply_patch when true. |
MAX_FILE_SIZE_BYTES |
max_file_size_bytes |
1048576 |
Range: 1 to 100000000. Applies to text file reads/search inputs. |
MAX_COMMAND_SECONDS |
max_command_seconds |
30 |
Range: 1 to 600. Applies to subprocess execution. |
MAX_OUTPUT_BYTES |
max_output_bytes |
1048576 |
Range: 1 to 100000000. Applies to command and artifact output. |
ALLOWED_COMMANDS |
allowed_commands |
empty | Comma-separated executables such as python,pytest, or JSON objects with name, executable, and optional default_args. Executables must be bare names. |
TEST_COMMANDS |
test_commands |
empty | JSON list of objects with name and command, for example [{"name":"unit","command":["pytest","tests/unit"]}]. |
ARTIFACT_DIRECTORY |
artifact_directory |
artifacts under workspace root |
Must exist and be a directory. Artifact collection is category-bounded inside it. |
SECRET_PATTERNS |
secret_patterns |
one token/password/API-key regex | JSON list of regex strings or comma-separated regex strings. |
LOG_LEVEL |
log_level |
INFO |
One of DEBUG, INFO, WARNING, ERROR, CRITICAL. |
Example TOML:
[workbench_mcp]
workspace_root = "."
read_only_mode = true
max_file_size_bytes = 1048576
max_command_seconds = 30
max_output_bytes = 1048576
allowed_commands = [{ name = "python", executable = "python" }]
test_commands = [{ name = "smoke", command = ["python", "-c", "print('ok')"] }]
artifact_directory = "artifacts"
secret_patterns = ["(?i)(api[_-]?key|token|secret|password)\\s*[:=]\\s*[^\\s]+"]
log_level = "INFO"
Verified Stdio Quick Start
For an MCP client, configure stdio with:
{
"command": "uv",
"args": ["run", "workbench-mcp", "--transport", "stdio"],
"env": {
"WORKSPACE_ROOT": "/path/to/workspace",
"ARTIFACT_DIRECTORY": "/path/to/workspace/artifacts",
"READ_ONLY_MODE": "true"
}
}
Local stdio verification command:
uv run pytest tests/e2e/test_mcp_stdio.py
Do not configure HTTP ports for this project; HTTP is not implemented or tested here.
Verified Docker Quick Start
Build and smoke-test the local image:
docker build -t workbench-mcp:local .
uv run python scripts/run-container-smoke.py --image workbench-mcp:local
Compose smoke workflow:
mkdir -p .workbench-demo/workspace .workbench-demo/artifacts
printf 'hello compose\n' > .workbench-demo/workspace/smoke.txt
docker compose config
docker compose run --rm --build workbench-mcp-smoke
PowerShell equivalent for the mount preparation:
New-Item -ItemType Directory -Force .workbench-demo\workspace, .workbench-demo\artifacts | Out-Null
Set-Content -LiteralPath .workbench-demo\workspace\smoke.txt -Value "hello compose"
The container defaults to stdio, non-root UID/GID 10001:10001, no published ports, no
host networking, no Docker socket mount, dropped Linux capabilities in Compose, and
no-new-privileges:true in smoke workflows.
Demo
Run the public reproducible demo:
uv run python scripts/run-demo.py
The demo creates a temporary Git workspace, writes deterministic files, configures the real
server, invokes the MCP tools through a FastMCP client, runs an approved test, applies one
controlled patch, demonstrates a blocked traversal attempt, writes
artifacts/demo/workbench-demo-report.json, prints a short summary, and cleans up the
temporary workspace.
Testing Commands
uv run ruff format --check .
uv run ruff check .
uv run mypy src
uv run pytest -rs
uv run pytest --cov=workbench_mcp --cov-report=term-missing
uv build
uv run pytest tests/e2e/test_mcp_stdio.py
uv run python scripts/run-demo.py
Docker checks:
docker build -t workbench-mcp:local .
uv run python scripts/run-container-smoke.py --image workbench-mcp:local
docker compose config
docker compose run --rm --build workbench-mcp-smoke
Security Model
- Deny by default for commands and writes.
- Workspace paths are resolved with
pathliband must remain insideWORKSPACE_ROOT. - Artifact paths are resolved inside category directories under
ARTIFACT_DIRECTORY. - Symlink escapes, parent traversal, absolute-path escapes, oversized files, binary text reads, shell operators, executable paths, command timeouts, bounded output capture, and truncation are covered by source checks and tests.
- Expected service errors are converted to sanitized MCP
ToolErrormessages. - Git behavior is read-only, bounded by the configured command timeout/output limit, and disables external diff execution for diff-stat inspection.
- Docker smoke workflows use a non-root runtime user and avoid privileged mounts.
This is not a complete arbitrary-code execution sandbox. Allowlisting a powerful executable
such as python, pytest, or git still grants that executable whatever behavior it can
perform inside the configured workspace and OS permissions.
More detail: docs/threat-model.md and SECURITY.md.
Known Limitations
- Only stdio transport is implemented and verified.
- HTTP and Streamable HTTP are not configured, exposed, or tested by this project.
- The server relies on host OS permissions; it is not a VM, kernel sandbox, or container escape prevention system.
- Windows symlink security tests skip when the OS denies symlink creation with
WinError 1314; Linux CI runs those tests explicitly. - Command safety depends on narrow allowlists. Do not allowlist broad interpreters for untrusted workspaces unless the surrounding environment is disposable.
- Artifact collection supports approved text suffixes only.
- Docker and Compose smoke workflows are finite verification commands, not a long-running hosted service.
Troubleshooting
uvnot found: install uv or use the fulluv.exepath on Windows as shown above.workspace_root does not exist: setWORKSPACE_ROOTto an existing directory.artifact_directory does not exist: create the directory or setARTIFACT_DIRECTORY.writes are blocked because READ_ONLY_MODE is enabled: setREAD_ONLY_MODE=falseonly for workspaces where controlled patching is acceptable.executable is not allowlisted: add a bare executable name toALLOWED_COMMANDS.- Docker bind-mount permission failures on Linux: run smoke workflows with a host-compatible UID/GID, as documented in docs/runbook.md.
- HTTP port questions: there is no project HTTP listener in this release.
Evidence And Docs
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。