mcpField
S
siGit

siGit

Remote
sigit.siVersion Control6Updated today

About siGit

siGit is a sigit.si command-line client for developers and AI agents, with an MCP mode for repository work. It reuses CLI credentials and can infer the target repository from the current Git checkout.

Key features

  1. 01Browse and create repositories
  2. 02Manage PRs and issues
  3. 03Search code
  4. 04Expose MCP tools over stdio
  5. 05Filter toolsets or use read-only mode

Use cases

  1. 01Ask an agent to list open PRs in the current repository
  2. 02Investigate issues and search code from the terminal
  3. 03Restrict an agent to read-only repository tools

MCP support

`si mcp` exposes repository, issue, PR, and code commands as MCP tools over stdio, reusing the token stored by `si auth login`. sigit.si also offers a remote MCP server reachable over HTTP.

Configuration

Remote servers need no local install. Pick your client, then copy the config or install in one click.

Endpointhttps://sigit.si/api/v1/mcp
Copy install command
Claude CodeTerminal
Cursor.cursor/mcp.json
VS Code.vscode/mcp.json
Claude Desktopclaude_desktop_config.json
Windsurf~/.codeium/windsurf/mcp_config.json

Run commands in a terminal; paste JSON into the client’s MCP config and restart. Servers that need sign-in will prompt on first connect.

Tools

Details

From the project README

sigit-si-cli

si is the command line interface for sigit.si — sign in, manage repositories, review pull requests, triage issues, search code, and drive Cloud Sessions, all from the terminal. It's the gh of sigit.si.

Install

cargo install --path crates/cli

This installs a binary named si.

Usage

si auth login                     # sign in with your sigit.si account
si repo list                      # your repositories
si repo create my-project --private

# Inside a repo checkout, si infers OWNER/NAME from the sigit.si git remote —
# --repo is only needed when you're not standing in one.
si pr list
si pr view 42
si pr create --title "Fix the thing" --base main

si issue list --state open
si issue create --title "Bug: ..." --body "Steps to reproduce..."

si code search "TODO" --ref main
si code view README.md

si session new --title "Refactor auth"
si billing show

si browse                          # open the current repo on sigit.si
si api repos                       # raw escape hatch for any endpoint

Pass --json to any read command for machine-readable output.

As an MCP server

si mcp serves these same commands to an AI agent as MCP tools over stdio:

{
  "mcpServers": {
    "sigit": { "command": "si", "args": ["mcp"] }
  }
}

There is nothing to authenticate — the server reuses the token si auth login stored. Point your editor at the directory of a checkout and repository tools default to the repo you're standing in, so the agent can say "list the open pull requests" without naming one; create_pull_request likewise defaults its head to the checked-out branch.

Trim the tool list when an agent doesn't need all of it:

si mcp --read-only                    # drop every tool that writes
si mcp --toolsets repo,issue,pr,code  # or SIGIT_TOOLSETS=repo,issue

Toolsets are account, repo, issue, pr, code, hook, session, and billing. sigit.si also hosts a remote MCP server at /api/v1/mcp for clients that would rather connect over HTTP than spawn a binary.

Environments
si --environment dev repo list          # a local sigit.si server (localhost:3000)
SIGIT_HOST=staging.example.com si repo list

Credentials are stored per-environment in ~/.sigit/token (or ~/.sigit-dev/token). SIGIT_TOKEN overrides the stored token, which is how CI authenticates without an interactive login.

Workspace layout

  • crates/sigit-si-api — a standalone, typed Rust client for the sigit.si JSON API. No CLI dependencies (clap, dialoguer, …); usable from any Rust project that wants to talk to sigit.si, including the desktop app.
  • crates/cli — the si binary: command parsing, terminal rendering, git remote resolution, and credential storage, built on top of sigit-si-api.

Issues, pull requests, and code search have no REST surface on sigit.si yet — they're served over the same MCP (Model Context Protocol) JSON-RPC endpoint the AI agent uses. sigit-si-api speaks that transport internally; it's not visible at the call site.

MCP Registry listings

This repo publishes two entries to the official MCP Registry, one per way of reaching the same tools, so registry-aware clients can add either in one click:

ListingFileWhat it is
si.sigit/cliserver.jsonThis CLI: a cargo package listing that runs si mcp over stdio.
si.sigit/sigitserver-sigit.jsonThe hosted server at sigit.si/api/v1/mcp, a remote Streamable-HTTP listing. The endpoint sigit-si-api talks to, and the official server siGit Code bakes in.

The CLI holds server.json because it's what this repo ships, and that's the filename mcp-publisher defaults to. Sibling listings are named after their own leaf, so the file says which entry it publishes.

Both go out through the release-mcp-registry.yml workflow, and both prove namespace ownership with a DNS TXT record on sigit.si rather than GitHub OIDC, so one credential covers the pair. A package listing has a second hurdle: the registry fetches the crate from crates.io and looks for the mcp-name: si.sigit/cli marker in its README, which is why that marker sits in this file. Crate versions are immutable, so a release that ships without the marker can never be listed, and fixing it costs a new version. The internal setup and release runbook live with the server (private sigit-si repo).

Development

make build   # cargo build --workspace
make test    # cargo test --workspace
make lint    # cargo clippy --workspace --all-targets -- -D warnings
make fmt     # cargo fmt --all
make check   # fmt + lint + test

Disclaimer

All em dashes are written by Seto Elkahfi.

License

Apache-2.0

Comments

Markdown supported

New to MCP?

From what the protocol is to installing your first server, these answers cover most first questions.

An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.