Provides AI coding agents with sandboxed, policy-gated access to local repositories for searching, reading, patching, and running curated commands.
Also known as bgorzelic/ghostlink, @bgorzelic/ghostlink
- Stars
- 0
- Tools
- —
- Downloads, 30 days
- —
- Latest version
- 0.1.0
- Last commit
- Oct 5, 2026
- License
- ISC
What is GhostLink MCP?
GhostLink provides AI coding agents with sandboxed, policy-gated access to local repositories for searching, reading, patching, and running curated commands.
Summary generated from the project's README
About GhostLink MCP server
GhostLink provides a security-hardened environment for AI coding agents to interact with local codebases. It functions as a local-first Model Context Protocol server that uses STDIO transport to connect with clients like Claude Code, Claude Desktop, Cursor, or Windsurf. By setting a specific repository root via environment variables, developers can grant agents controlled access to search, read, and modify files without exposing the entire system to arbitrary shell commands.
The server includes six specialized tools designed for the autonomous coding cycle. These include repo.search for regex searches using ripgrep, repo.read_file for reading content with size limits, and repo.apply_patch for making code changes with atomic rollback capabilities. It also offers repo.run for executing a specific allowlist of commands like test, lint, or build, alongside git.status and git.diff for version control tasks. Every interaction is sandboxed to the designated directory and returns a deterministic JSON envelope for predictable agent behavior.
What you can do with it
- 1Auditing agent actionsTeams can maintain a JSONL audit trail of every tool invocation to monitor how an AI interacts with the code.
Written from the project's README. Every use case is checked against a quote from it.
How to install GhostLink MCP
Formats verified 10 Nov 2026Add to CursorAdd to GooseInstall GhostLink MCP in Claude Code
$ claude mcp add --env 'GHOSTLINK_REPO_ROOT=<GHOSTLINK_REPO_ROOT>' ghostlink -- npx -y @bgorzelic/ghostlinkInstall GhostLink MCP in Cursor
~/.cursor/mcp.json
{
"mcpServers": {
"ghostlink": {
"command": "npx",
"args": [
"-y",
"@bgorzelic/ghostlink"
],
"env": {
"GHOSTLINK_REPO_ROOT": "<GHOSTLINK_REPO_ROOT>"
}
}
}
}Install GhostLink MCP in VS Code
.vscode/mcp.json
{
"servers": {
"ghostlink": {
"type": "stdio",
"command": "npx",
"args": [
"-y",
"@bgorzelic/ghostlink"
],
"env": {
"GHOSTLINK_REPO_ROOT": "<GHOSTLINK_REPO_ROOT>"
}
}
}
}Install GhostLink MCP in VS Code CLI
$ code --add-mcp '{"name":"ghostlink","type":"stdio","command":"npx","args":["-y","@bgorzelic/ghostlink"],"env":{"GHOSTLINK_REPO_ROOT":"<GHOSTLINK_REPO_ROOT>"}}'Install GhostLink MCP in Claude Desktop
claude_desktop_config.json
{
"mcpServers": {
"ghostlink": {
"command": "npx",
"args": [
"-y",
"@bgorzelic/ghostlink"
],
"env": {
"GHOSTLINK_REPO_ROOT": "<GHOSTLINK_REPO_ROOT>"
}
}
}
}Install GhostLink MCP in Codex
$ codex mcp add ghostlink --env 'GHOSTLINK_REPO_ROOT=<GHOSTLINK_REPO_ROOT>' -- npx -y @bgorzelic/ghostlinkInstall GhostLink MCP in Codex config.toml
~/.codex/config.toml
[mcp_servers.ghostlink] command = "npx" args = ["-y", "@bgorzelic/ghostlink"] [mcp_servers.ghostlink.env] GHOSTLINK_REPO_ROOT = "<GHOSTLINK_REPO_ROOT>"
Install GhostLink MCP in Gemini CLI
$ gemini mcp add -s user -e 'GHOSTLINK_REPO_ROOT=<GHOSTLINK_REPO_ROOT>' ghostlink npx -- -y @bgorzelic/ghostlinkInstall GhostLink MCP in Gemini settings.json
~/.gemini/settings.json
{
"mcpServers": {
"ghostlink": {
"command": "npx",
"args": [
"-y",
"@bgorzelic/ghostlink"
],
"env": {
"GHOSTLINK_REPO_ROOT": "<GHOSTLINK_REPO_ROOT>"
}
}
}
}Configuration
GHOSTLINK_REPO_ROOTRequiredAbsolute path to the repository the server is confined to. Every tool call is sandboxed beneath this path.
GHOSTLINK_LOGOptional path for the JSON audit log of tool calls. Omit to disable file logging.
| Name | Description | |
|---|---|---|
| GHOSTLINK_REPO_ROOT | Absolute path to the repository the server is confined to. Every tool call is sandboxed beneath this path. | Required |
| GHOSTLINK_LOG | Optional path for the JSON audit log of tool calls. Omit to disable file logging. |
Automated checks
Ran Nov 14, 2026Full report →Passed our automated checks. That is not a security audit: read the code and give the server the narrowest access it needs. How checks work
- Source repository linked and reachablebgorzelic/ghostlink is public and reachablegithub.com/bgorzelic/ghostlink ↗Pass
- Package names this repository as its sourceNot checked
- Build provenance (Sigstore attestation)Not checked
- Open source license (OSI)ISCspdx.org/licenses/ISC.html ↗Pass
- Commit or release in the last yearLast commit or release 2026-10-05github.com/bgorzelic/ghostlink ↗Pass
- OpenSSF ScorecardNot checked
- Package name not close to a popular packageNo popular package one edit away from @bgorzelic/ghostlinkPass
- No critical advisory in the package or its dependenciesthe package checked; no critical advisoryosv.dev ↗Pass
- No hidden instructions to the model0 tool descriptions and the README read; no override phrases, credential paths or hidden charactersPass
- Destructive tools are labelledNot checked
- Remote endpoint answers the MCP handshakeNot checked
README
View on GitHubGhostLink
A security-hardened MCP server that gives AI coding agents safe, deterministic access to local repositories.
Add it to any MCP client that supports STDIO. For Claude Code, create .mcp.json in the target repo root:
{
"mcpServers": {
"ghostlink": {
"command": "npx",
"args": ["-y", "ghostlink"],
"env": {
"GHOSTLINK_REPO_ROOT": "/path/to/target/repo"
}
}
}
}
Then run claude in that directory — six repo tools appear, all confined to GHOSTLINK_REPO_ROOT. Every tool call returns the same deterministic ToolEnvelope:
{
"ok": true,
"data": { ... },
"provenance": { "tool": "repo.search", "timestamp": "2026-02-24T...", "duration_ms": 42 }
}
On error, "error": { "code": "...", "message": "..." } replaces "data". Full tool schemas: docs/TOOLS.md.
Tools
| Tool | Description |
|---|---|
repo.search |
Ripgrep-powered regex search with glob filtering, deterministic ordering, and output caps (max 200 results) |
repo.read_file |
File read with size caps (max 10MB), binary detection, and truncation flags |
repo.apply_patch |
Unified diff patching with dry-run mode, full sandbox validation, and atomic rollback on failure |
repo.run |
Curated command execution (test, lint, typecheck, build, smoke) -- no arbitrary shell, allowlisted args only |
git.status |
Normalized git status with branch info, ahead/behind tracking, and sorted file entries |
git.diff |
Staged or unstaged diff with path filtering, sandbox validation, and output caps (max 2MB) |
What is GhostLink?
GhostLink is a local-first Model Context Protocol server that exposes your codebase to AI coding agents through a small set of policy-gated tools. It solves a specific problem: AI agents need to search, read, patch, and verify code, but giving them raw shell access is a liability. GhostLink provides a sandboxed capability plane where every tool call is confined to a single repository root, every output follows a deterministic JSON shape, and every invocation is audit-logged.
Why GhostLink?
| Capability | What it means |
|---|---|
| Secure local dev plane | Repo-root sandbox, no shell execution, JSONL audit trail on every tool call |
| Deterministic output | Same input produces the same JSON envelope shape -- enables golden tests and predictable agent consumption |
| Policy enforcement | Command allowlists, output caps, truncation flags, timeout enforcement -- the AI cannot do unbounded damage |
| Agent loop foundation | Built for the search, read, patch, verify cycle that autonomous coding agents run in a loop |
| Multi-server composition | One GhostLink instance per repo, composable with other MCP servers in the same client session |
| Production-ready Phase 2 base | Transport abstraction, schema versioning, and auth hook seams are preserved in the architecture today |
Architecture
GhostLink is a three-layer stack designed for extensibility without core changes:
flowchart TD
T["Transport -- src/index.ts<br/>STDIO now, HTTP/SSE in Phase 2"]
S["Server factory -- src/server.ts<br/>Transport-agnostic tool registration via MCP SDK + Zod schemas"]
TL["Tools -- src/core/tools/*<br/>Six tools, each returning ToolEnvelope<T>"]
P["Policy -- src/core/policy/*<br/>Sandbox enforcement, audit logging, output caps"]
T --> S --> TL --> P
The createServer() factory knows nothing about transport. Adding HTTP/SSE in Phase 2 means writing a new transport binding and auth middleware -- the server factory and all tool implementations remain unchanged. Phase 3 (agent runtime) adds memory resources and orchestration as consumers of GhostLink, not modifications to it.
Quick Start
Prerequisites
- Node.js 18+
- ripgrep (
brew install ripgrep) - A git repository to expose
Install
From npm:
npm install ghostlink
Or from source:
git clone https://github.com/bgorzelic/ghostlink.git
cd ghostlink
npm install
npm run build
Smoke Test (Raw STDIO)
GhostLink speaks JSON-RPC 2.0 over STDIO. Test it directly:
echo '{"jsonrpc":"2.0","id":1,"method":"tools/list"}' | \
GHOSTLINK_REPO_ROOT=/path/to/target/repo node dist/index.js
This returns all 6 tools and their schemas.
Client Configuration
GhostLink works with any MCP client that supports STDIO transport. The npx snippet at the top of this page works everywhere; a source checkout uses node with the built entry point instead:
{
"ghostlink": {
"command": "node",
"args": ["/absolute/path/to/ghostlink/dist/index.js"],
"env": {
"GHOSTLINK_REPO_ROOT": "/path/to/target/repo"
}
}
}
| Client | Where the config goes |
|---|---|
| Claude Code | .mcp.json in the target repo root (mcpServers key), then run claude there |
| Claude Desktop | ~/Library/Application Support/Claude/claude_desktop_config.json (mcpServers key), then restart |
| Cursor, Windsurf, Cline, others | Your client's MCP server configuration -- consult its documentation for the file location |
The transport is always STDIO. Ready-to-use .mcp.json and CLAUDE.md templates for target projects live in templates/.
Security Model
GhostLink enforces defense-in-depth at every layer:
- Repo-root sandbox -- All file operations confined to
GHOSTLINK_REPO_ROOT. Path traversal, symlink escape, null bytes, and absolute paths outside the root are all rejected before any filesystem access. - No shell execution --
repo.runusesspawnwithshell: false. Commands are limited to a fixed allowlist (test, lint, typecheck, build, smoke) with per-command argument allowlists. Environment is stripped to six safe variables. - Output caps -- Every tool that returns bulk data enforces hard maximums (200 search results, 10MB file reads, 200KB stdout/stderr, 2MB diffs). Truncation is flagged, never silent.
- Atomic patch rollback --
repo.apply_patchvalidates all paths and computes all patches before writing anything. If any write fails, completed writes are rolled back to their original state. - Timeout enforcement --
repo.runkills processes at configurable timeouts (default 120s, hard cap 300s) with SIGTERM then SIGKILL.
Full threat model and mitigations: docs/SECURITY.md.
Audit Logging
Every tool call produces a JSONL audit entry: {ts, tool, ok, duration_ms, error_code?, repo_root}.
GHOSTLINK_LOG |
Behavior |
|---|---|
stdout (default) |
JSONL audit lines written to stderr |
file |
JSONL written to logs/ghostlink.jsonl (auto-rotates at 10MB) |
off |
No logging |
Set via environment variable:
GHOSTLINK_LOG=file GHOSTLINK_REPO_ROOT=/path/to/repo node dist/index.js
Prompt Templates
docs/PROMPTS.md contains ready-to-use prompts for high-autonomy agent operation, including orchestrator prompts, sub-agent role definitions (Protocol Engineer, Toolsmith, Security Reviewer, Test Engineer, Docs Engineer), and multi-instance coordination patterns.
Development
npm install # Install dependencies
npm test # Run test suite (108 tests via Vitest)
npm run lint # ESLint
npm run typecheck # TypeScript strict mode check
npm run build # Compile to dist/
npm run dev # Dev mode with auto-reload (tsx watch)
Full verification after edits:
npm test && npm run lint && npm run typecheck && npm run build
Documentation
| Document | Description |
|---|---|
| docs/TOOLS.md | Canonical tool schemas (versioned public API) |
| docs/SECURITY.md | Threat model and mitigations |
| docs/QUICKSTART.md | Setup, smoke tests, and client configuration walkthrough |
| docs/INSPECTOR.md | MCP Inspector manual testing guide |
| docs/PROMPTS.md | Agent prompts for orchestration and sub-agent roles |
| docs/ROADMAP_DETAILED.md | Full product roadmap with Phase 2 and Phase 3 deliverables |
| docs/WHY_GHOSTLINK.md | Strategic value proposition and architecture rationale |
| docs/ENGINEERING_REPORT_v0.1.0.md | v0.1.0 ship report with milestone history and decision log |
| templates/ | Ready-to-use CLAUDE.md and .mcp.json templates for target projects |
Roadmap
Phase 1 -- Local STDIO [Shipped, v0.1.0]
Deterministic tool surface, repo-root sandbox, curated command execution, 108 tests, JSONL audit logging, npm package published.
Phase 2 -- Remote Transport [Planned]
HTTP/SSE transport, OAuth 2.1 authentication, multi-user tenant separation, per-tenant rate limiting, schema versioning, structured audit logging with correlation IDs.
Phase 3 -- Agent Runtime [Future]
Persistent memory resources exposed via MCP, optional policy-gated memory write tools, orchestration layer (external to GhostLink), evaluation loops, sub-agent coordination framework.
License
ISC
GhostLink MCP questions
How do I add GhostLink MCP to Claude Code?
Run claude mcp add --env 'GHOSTLINK_REPO_ROOT=<GHOSTLINK_REPO_ROOT>' ghostlink -- npx -y @bgorzelic/ghostlink in your terminal. Add --scope user to make it available in every project.
How do I add GhostLink MCP to Cursor?
Open ~/.cursor/mcp.json and add the server under mcpServers, using the Cursor config shown in the install section above.
How do I configure GhostLink for Claude Code?
Create a .mcp.json file in your target repository root containing the ghostlink command and the GHOSTLINK_REPO_ROOT environment variable.
Is it safe to give an AI agent access to my files with GhostLink?
Yes, it uses a repo-root sandbox that rejects path traversal and symlink escapes, and it prevents arbitrary shell execution by using an allowlist of commands.
What happens if a file patch fails?
The server performs an atomic rollback to return the files to their original state if any part of the write operation fails.
Can I see a log of what the AI agent has done?
Yes, every tool call produces a JSONL audit entry that can be sent to stdout or a file.