ai-git
A deterministic MCP server that reimplements ai-git-fish workflows (aicommit, aibranch, aipr) with Leantime integration, enabling Claude Code and Codex to perform git operations and manage tickets without an internal LLM.
README
ai-git-mcp
An MCP server that reimplements the ai-git-fish
workflows — aicommit, aibranch, aipr — with Leantime integration, without fish shell and
without DeepSeek. Wire it up once and the same workflows work in Claude Code and Codex.
- Why no AI inside the server?
- Requirements
- Installation
- Configuration (env)
- Usage — slash commands
- Tools reference
- Troubleshooting
- Local development
Why no AI inside the server?
ai-git-fish calls DeepSeek to generate commit messages and PR text. But Claude Code and Codex
are already the AI. So this server contains no LLM: every tool is deterministic. The agent
reads the diff/ticket through a tool, writes the prose itself, and passes it into the next tool.
No paying for AI twice, no losing context, and the agent understands your code better than DeepSeek.
Agent (Claude/Codex) ──calls tools──> ai-git-mcp ──> git / gh CLI
│ │
└── writes commit/PR/slug └──> Leantime JSON-RPC (/api/jsonrpc)
Requirements
On the machine that runs the agent:
- Node.js ≥ 18 (
node -v) - git in
PATH(git --version) - GitHub CLI
gh, authenticated (gh auth login) — needed only foraipr - A Leantime instance + API key — needed only for the ticket tools
You do not need to clone this repo. npx fetches and builds it for you (see below).
Installation
1. Get your Leantime credentials
| Value | Where to find it |
|---|---|
LEANTIME_BASE_URL |
Your Leantime URL, e.g. https://leantime.example.com |
LEANTIME_API_KEY |
Leantime → your profile → API Keys → generate one |
LEANTIME_USER_ID |
Your numeric user id — see the tip below |
Finding your
LEANTIME_USER_IDthe easy way (replaces ai-git-fish's installer fzf picker): set justLEANTIME_BASE_URL+LEANTIME_API_KEYfirst, restart, then ask the agent to runleantime_list_users. It prints every user with theirid— copy yours into the env and restart. (Or: open your profile and read the id from the URL, e.g..../users/editUser/5→5.)
2. Add the server to your client
Pick your client. The npx -y github:BenBen2109/ai-git-mcp command downloads this repo, builds
it once, caches it, and runs it — no manual clone or build.
<details open> <summary><b>Claude Code</b></summary>
Option A — one command:
claude mcp add ai-git \
-e LEANTIME_BASE_URL=https://leantime.example.com \
-e LEANTIME_API_KEY=your-key \
-e LEANTIME_USER_ID=123 \
-- npx -y github:BenBen2109/ai-git-mcp
Option B — edit .mcp.json (project root) or ~/.claude.json (global):
{
"mcpServers": {
"ai-git": {
"command": "npx",
"args": ["-y", "github:BenBen2109/ai-git-mcp"],
"env": {
"LEANTIME_BASE_URL": "https://leantime.example.com",
"LEANTIME_API_KEY": "your-key",
"LEANTIME_USER_ID": "123"
}
}
}
}
</details>
<details open> <summary><b>Codex</b></summary>
Edit ~/.codex/config.toml:
[mcp_servers.ai-git]
command = "npx"
args = ["-y", "github:BenBen2109/ai-git-mcp"]
env = { LEANTIME_BASE_URL = "https://leantime.example.com", LEANTIME_API_KEY = "your-key", LEANTIME_USER_ID = "123" }
</details>
Pin a version by appending a tag/branch/commit:
github:BenBen2109/ai-git-mcp#v1.0.0. The first launch is slower (it builds once); later launches use the npx cache.
3. Restart and verify
Restart the client (or reload the MCP config). Then check the server is connected:
- Claude Code: run
/mcp—ai-gitshould show 14 tools and 4 prompts. - Codex: the
ai-gittools appear in the tool list at startup.
Configuration (env)
| Variable | Required | Meaning |
|---|---|---|
LEANTIME_BASE_URL |
ticket tools | e.g. https://leantime.example.com |
LEANTIME_API_KEY |
ticket tools | sent as the x-api-key header |
LEANTIME_USER_ID |
recommended | numeric id used to filter your tickets |
LEANTIME_PICK_LIMIT |
optional | max tickets listed (default 30) |
SLACK_WEBHOOK_URL |
Slack tools | Incoming Webhook URL; the channel is fixed on the Slack side |
PR_PREVIEW_URL_TEMPLATE |
optional | Deploy-preview URL with a {pr} placeholder, e.g. https://frontend-preview-{pr}--anymatemeproject.netlify.app. Adds a Preview line to the Slack PR message |
SLACK_USER_MAP |
optional | JSON map of name/login → Slack member ID, e.g. {"khanhld":"U012ABC"}, for @-mentions |
SLACK_PR_REVIEWERS |
optional | Who to @-mention on every PR — comma list of names (from the map), raw Slack ids, or here/channel |
The git-only tools (git_status, git_staged_diff, git_commit, git_log_for_pr,
git_create_branch, git_create_or_update_pr, git_undo_last_commit) work without any Leantime
config. See .env.example.
Usage — slash commands
Three workflows, exposed as MCP prompts. In Claude Code type the slash command; in Codex just say it in plain language (e.g. "run aicommit") — the agent calls the same tools either way.
/ai-git:aicommit [hint] — stage-aware commit
Reads your staged diff, writes a Conventional Commit message, shows it to you, then commits.
# stage what you want first
git add -p
# then, in the agent:
/ai-git:aicommit
/ai-git:aicommit drop the debug logging # optional hint to steer the message
→ git_staged_diff → (agent writes message) → git_commit
/ai-git:aibranch [type] — branch from a Leantime ticket
Lists your tickets, you pick one, it creates/switches to type/<id>-<slug> (slug auto-derived
from the ticket headline, Vietnamese diacritics handled).
/ai-git:aibranch # type defaults to "feat" → feat/1234-add-login
/ai-git:aibranch fix # → fix/1234-add-login
→ leantime_list_my_tickets → (you pick) → git_create_branch
If LEANTIME_USER_ID isn't set (or you want someone else's tickets), the agent calls
leantime_list_users first, you pick a person, and it lists that user's tickets via assignee_id
— the equivalent of ai-git-fish's fzf user picker.
/ai-git:aipr [hint] — open or update a GitHub PR
Pushes the branch and creates (or updates) its PR with an AI-written title and body. If the branch
is named type/<id>-<slug>, the Leantime ticket is auto-linked and embedded.
/ai-git:aipr
/ai-git:aipr focus on the migration # optional hint to steer the PR
→ git_status + git_log_for_pr → (agent writes title/body) → git_create_or_update_pr → slack_post_pr
After the PR is created/updated, it sends a formatted notification to the team Slack channel
(via SLACK_WEBHOOK_URL). If Slack isn't configured, that step is skipped with a note.
/ai-git:aireview <pr> [spec] — review a PR and post the review
Fetches the PR + diff, reviews it against the team checklist (Functionality, Security, Error Handling, Code Quality, Performance, Style), writes the review in English following a fixed format, and posts it as a PR comment. Optionally validates against a Redmine spec URL.
/ai-git:aireview https://github.com/org/repo/pull/42
/ai-git:aireview https://github.com/org/repo/pull/42 https://redmine.example.com/issues/1234
Prefer the full PR URL — it tells gh which repo the PR is in. A bare number (42) only
resolves when repo_path is the PR's repository.
→ git_pr_view → (agent reviews, + Redmine spec via the agent's redmine tools if given) → git_pr_comment
Boundaries: never approves or merges, never changes code without approval.
Tools reference
All tools return JSON. Each accepts an optional repo_path (defaults to the server's working
directory) unless noted.
Leantime
| Tool | Params | Returns |
|---|---|---|
leantime_list_users |
— | { count, users: [{ id, name, username, isConfigured }] } (setup helper to find your LEANTIME_USER_ID) |
leantime_list_my_tickets |
limit? (1–100), assignee_id? |
{ count, assigneeId, filtered, tickets: [{ id, headline, statusLabel, projectName, assignee, dueDate, url, description }] } (description is a ~200-char snippet, like the fzf preview; filtered:false means no LEANTIME_USER_ID and no assignee_id → all tickets) |
leantime_get_ticket |
id |
{ found, id, headline, description, statusLabel, dueDate, url, markdown } |
Git — read-only
| Tool | Params | Returns |
|---|---|---|
git_status |
repo_path? |
{ branch, ahead, behind, staged[], unstaged[], untracked[] } |
git_staged_diff |
repo_path?, max_bytes? |
{ diff, files: [{ path, status }], isEmpty, truncated } |
git_log_for_pr |
base?, repo_path? |
{ base, count, commits: [{ sha, subject, body }] } |
git_pr_view |
pr, max_bytes?, repo_path? |
{ number, title, body, state, author, baseRefName, headRefName, url, additions, deletions, changedFiles, files[], diff, diffTruncated } |
Git — write
| Tool | Params | Returns |
|---|---|---|
git_commit |
message, repo_path? |
{ sha, shortStat } |
git_create_branch |
ticket_id, type?, slug?, repo_path? |
{ branch, created } |
git_create_or_update_pr |
title, body, ticket_id?, base?, draft?, repo_path? |
{ url, number, action } |
git_pr_comment |
pr, body, repo_path? |
{ url } — posts a comment (e.g. a review); never approves/merges |
git_undo_last_commit |
repo_path? |
{ undoneSha, message } |
Slack
| Tool | Params | Returns |
|---|---|---|
slack_post_pr |
title, url, number?, author?, ticket_id?, mention? |
{ ok } — posts Author / PR / Preview / Ticket / Feat / Reviewers (@-mentions) to the webhook channel |
slack_send_message |
text, mention? |
{ ok } — plain Slack mrkdwn message, optional @-mentions |
@-mentions: Slack only pings by member id (<@U…>), not by display name. Set SLACK_USER_MAP
to map names → ids, then list reviewers in SLACK_PR_REVIEWERS (or pass mention per call).
here/channel work without any map.
Notes:
git_commitfails with an actionable message if nothing is staged.git_create_branchauto-derives the slug from the ticket headline whenslugis omitted.git_create_or_update_prpushesHEAD, detects an existing PR for the branch (edits it) or creates a new one, and appends the ticket markdown block whenticket_idis given.git_undo_last_commitis a soft reset — your changes stay staged, nothing is lost.
Troubleshooting
| Symptom | Fix |
|---|---|
Leantime is not configured |
Set LEANTIME_BASE_URL + LEANTIME_API_KEY in the server env. |
Not a git repository |
Run the agent from inside a git repo, or pass repo_path. |
Nothing staged to commit |
git add your changes first. |
gh ... failed on aipr |
Run gh auth login; ensure the repo has a GitHub remote. |
npx is slow on first run |
Expected — it builds once, then caches. Pin a tag to stabilise. |
| Slash commands missing in Codex | Codex prompt support varies; just say "run aicommit" instead. |
Local development
git clone https://github.com/BenBen2109/ai-git-mcp
cd ai-git-mcp
npm install
npm run build # → dist/index.js
npm test # vitest
npm run dev # tsx watch (live reload)
Local-path config (instead of npx):
{ "mcpServers": { "ai-git": {
"command": "node",
"args": ["/absolute/path/to/ai-git-mcp/dist/index.js"],
"env": { "LEANTIME_BASE_URL": "...", "LEANTIME_API_KEY": "...", "LEANTIME_USER_ID": "..." }
}}}
Credits
Leantime client and ticket formatting (HTML/entity decoding, status labels, PR markdown block)
are ported from ai-git-fish by DucTam2411.
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.