git-ai

LLM-powered git workflow tools. Generate Conventional Commits messages and PR descriptions from your staged changes and branch — in the CLI, Lazygit, or any git environment.

Install

npm install -g @waxmard/git-ai

Or clone and symlink for local development (edits are live):

make install   # symlinks to ~/.local/bin and ~/.local/lib
make uninstall

See CONTRIBUTING.md for the full dev setup, test commands, and PR/release process.

Quickstart

Run the setup wizard once — it detects your installed provider CLIs and keys, helps you authenticate, and writes your config:

git-ai setup

If you already have a provider CLI installed or a key in your environment, the wizard does the work for you: it enables every provider you can use right now, pins a sensible recommended model for each, and drops you on a summary of the resulting config — a single Enter finishes, or pick an action there to tweak it (add/remove providers, change models, reset and re-detect). No provider/model picking required. Set GIT_AI_NO_SETUP_FAST=1 to skip straight to the full manual picker.

The first time you run git-ai commit or git-ai pr with nothing configured, the wizard launches automatically (skip with GIT_AI_NO_SETUP=1). Then generate:

git add -A
git-ai commit                         # prints a Conventional Commits message to stdout
git commit -m "$(git-ai commit)"      # …or commit with it in one line

git-ai pr --base main                 # PR title + body for the current branch

With no auth method on the command line, git-ai uses your configured default or pops an interactive fzf picker. You can still pass one explicitly — git-ai commit gemini-api. Both git-ai and aigit are interchangeable.

Auth methods

git-ai setup is the way in — it shows which of these are ready, authenticates you, and writes the config. The table is a reference for what exists; you don't configure any of it by hand unless you want to (see Manual configuration).

git-ai needs at least one of these. gemini-api (Gemini CLI) and the two Vertex methods are the common choices; the rest are available if you already use that provider's CLI or API.

| Auth Method | Runtime | Credentials | |-------------|---------|-------------| | gemini-api | Gemini CLI | GEMINI_API_KEY or system keychain | | vertex-gemini | curl + python3 + gcloud | Google ADC / Vertex credentials | | vertex-anthropic | curl + python3 + gcloud | Google ADC / Vertex credentials | | claude-code | Claude Code CLI | Claude Code CLI session | | codex | Codex CLI | Codex CLI session | | anthropic-api | curl + python3 | ANTHROPIC_API_KEY | | openai-api | curl + python3 | OPENAI_API_KEY |

curl and python3 are standard on macOS and most Linux systems.

Vertex AI support covers only the Gemini (vertex-gemini) and Anthropic (vertex-anthropic) model families. Other Vertex publishers (Meta Llama, Mistral, etc.) are not yet supported. To pin a GCP account or run multiple projects, see Pinning a GCP account (Vertex).

For API-key providers (gemini-api, anthropic-api, openai-api), git-ai setup prompts for the key and stores it in your OS keychain or shell rc. For Google ADC / service-account credentials, use a vertex-gemini or vertex-anthropic method and let setup run gcloud auth application-default login. To wire any of this up by hand instead, see Manual configuration.

Commands

setup

Interactive wizard to configure providers and authentication.

git-ai setup

• First run (no config yet): enables every provider that already authenticates, pins each one's recommended model, and drops you on the config overview — a bare Enter finishes. For Vertex AI it also picks your GCP project automatically (your active gcloud project if it has the Vertex API enabled, else the first of your projects that does). When nothing is ready (or GIT_AI_NO_SETUP_FAST=1 is set), falls back to a readiness table and a manual provider/model picker (each family's recommended model leads the list) that exits when done — re-run git-ai setup anytime to make changes
• Later runs (config exists): opens the overview with an edit menu — add a provider, remove one, change a provider's models, change Vertex AI projects, or reset (re-detect and start over) — applied in place, one change at a time, preserving everything else in the file (comments, vertex account=/projects= settings). Only a confirmed reset rewrites the file wholesale. The models and projects edits are both replace-style multi-selects: current entries are pre-listed, and the set you mark replaces the old one, so a single pass adds and removes (Esc or a blank entry keeps things as they are)
• Shows a single Vertex AI entry; whether a model runs via vertex-gemini or vertex-anthropic is inferred from the model id, never asked
• For API-key providers, prompts for the key and offers to store it in your OS keychain or shell rc
• For Vertex AI, offers to run gcloud auth application-default login and prompts for project (required) / region / account, written to the shared [vertex] block. Profiles and credentials= stay manual — see Pinning a GCP account (Vertex)
• Seeds the per-repo default so the next commit/pr runs without prompting
• Runs automatically on first use when nothing is configured; set GIT_AI_NO_SETUP=1 to disable that

commit

Generate a commit message from staged changes.

git-ai commit [auth-method] [model-id]

• Reads git diff --staged and produces a Conventional Commits message
• Includes a description body for non-trivial changes
• Uses your configured default (from git-ai setup) or the interactive picker; pass an auth method to override
• All auth methods default to a lightweight model when model-id is omitted
• Pass last as the provider to reuse the previously generated message

pr

Generate a PR title and body from the current branch.

git-ai pr [auth-method] [model-id] [--base <branch>] [--fresh] [--from-sha <commit>]
git-ai mr [...]   # alias for pr

• Reads the commit log and diff against the base branch
• Produces a Conventional Commits title + markdown body with a ### Test Plan section
• Auto-detects the base branch from the remote default (falls back to main)
• Use --base to override (e.g. --base dev)
• Saves the generated output per current-branch/base-branch pair under .git/pr-cache/; subsequent runs with the same pair refine the previous result automatically
• Use --fresh to ignore the saved output and regenerate from scratch
• Use --from-sha to override the saved HEAD and regenerate only from commits after a specific prior generated commit
• Uses your configured default (from git-ai setup) or the interactive picker; pass an auth method to override
• All auth methods default to a stronger model when model-id is omitted

options

List every auth-method / model combo as a flat pipe-delimited list, LRU-sorted. Primary input for the fzf-based Lazygit integration; also useful for custom pickers.

git-ai options [commit|pr]

• Emits one provider:model|<label> line per selectable combo
• For commit, also emits last|reuse saved message when a saved message exists
• Most-recent picks (from .git/{tool}-choice-history) float to the top; remaining combos follow in default order
• git-ai commit <provider:model> and git-ai pr <provider:model> accept the emitted value directly

providers / models

List available auth methods and models, ordered by last-used. Kept for scripting and as a fallback when options isn't a fit.

last is only a commit provider option; PR refinement reuses cached prior output automatically.

git-ai providers [commit|pr]
git-ai models <auth-method> [commit|pr]

Python library

git-ai is also distributed as a Python package (waxmard-git-ai) so other tools can reuse the same commit-message and MR-description prompt assembly without shelling out.

pip install waxmard-git-ai
# or: uv add waxmard-git-ai

Bring your own Claude / Gemini / OpenAI / ADK / anything — sync or async.

Commit message (data-mode):

import git_ai

system, user = git_ai.build_commit_prompt(diff_text)
raw = my_llm(system, user)               # your call: SDK, agent framework, REST, etc.
commit_msg = git_ai.parse_commit_response(raw)

MR/PR description (data-mode — no local checkout, e.g. fetched from the GitHub/GitLab API):

import git_ai

log = git_ai.format_commit_log((c.title, c.message) for c in mr_commits)
system, user = git_ai.build_mr_prompt(
    diff=diff_text,
    commit_log=log,
    existing_pr=current_pr_body or None,
)
raw = my_llm(system, user)
pr_text = git_ai.parse_mr_response(raw)

# Optional: render a compact ~ / + / - delta against the prior PR
delta = git_ai.render_pr_diff(current_pr_body, pr_text, color=False) or None

diff_stat and release_context are optional — when omitted, the diff-stat is derived from the diff and a generic "no release tags found" context is used. Model selection, retries, auth, and error handling are the caller's responsibility (inside my_llm).

Repo-mode (reads staged diff / base..HEAD from a local checkout):

import git_ai

# Commit message from staged changes (auto-loads .git-ai-ignore)
diff = git_ai.get_staged_diff(".")
system, user = git_ai.build_commit_prompt(
    diff, release_context=git_ai.get_release_context("."),
)
commit_msg = git_ai.parse_commit_response(my_llm(system, user))

# PR description with incremental cache reuse
ctx = git_ai.prepare_repo_pr_context(".", base_branch="main")
if ctx.no_changes:
    pr_text = ctx.existing_pr            # HEAD unchanged, reuse cached PR
else:
    system, user = git_ai.build_mr_prompt(
        diff=ctx.diff,
        commit_log=ctx.commit_log,
        diff_stat=ctx.diff_stat,
        release_context=ctx.release_context,
        existing_pr=ctx.existing_pr,
    )
    pr_text = git_ai.parse_mr_response(my_llm(system, user))
    if ctx.current_branch:
        git_ai.save_cached_pr(
            git_ai.get_git_dir("."),
            ctx.current_branch,
            "main",
            pr_text,
            ctx.head_sha,
        )

prepare_repo_pr_context reuses .git/pr-cache/ automatically, sets no_changes=True when HEAD matches the cached SHA (so callers can skip the LLM entirely), and narrows the diff/commit_log to commits after the last generated HEAD when possible. Pass fresh=True to bypass the cache for one call, or previous_head_sha= to override the cached incremental base explicitly.

Data-mode is stateless. To get the same efficiency in remote consumers, persist the prior PR text + last generated head SHA yourself, fetch the incremental diff/log since that SHA from your SCM, and pass them to build_mr_prompt(diff=..., commit_log=..., existing_pr=...).

Async / agent-framework example — the prompt builders are pure, so anything goes inside the LLM call. Pass system and user to whatever your SDK expects (Anthropic system= + messages=[{"role": "user", ...}], OpenAI/Gemini message lists, ADK agent instruction + input, etc.):

import git_ai
from anthropic import AsyncAnthropic

client = AsyncAnthropic()

async def commit_msg(diff: str) -> str:
    system, user = git_ai.build_commit_prompt(diff)
    resp = await client.messages.create(
        model="claude-sonnet-4-6",
        max_tokens=1024,
        system=system,
        messages=[{"role": "user", "content": user}],
    )
    return git_ai.parse_commit_response(resp.content[0].text)

Excluding noisy files (.git-ai-ignore)

Lockfiles and other generated artifacts can dominate a diff and push it past the LLM provider's input cap. git-ai always excludes the following filenames from git diff --staged (commit) and git diff base...HEAD (pr):

package-lock.json    yarn.lock         pnpm-lock.yaml      npm-shrinkwrap.json
Gemfile.lock         Cargo.lock        go.sum              poetry.lock
uv.lock              composer.lock     Pipfile.lock        pubspec.lock
mix.lock             flake.lock

Drop a .git-ai-ignore file at the repo root to add more patterns (one per line, # comments and blank lines ignored). Patterns are Git pathspec glob fragments that git-ai prefixes with **/, so generated/**/*.ts matches TypeScript files under any generated/ directory; leading / is not .gitignore root syntax. Lines starting with ! re-include a pattern, useful when you actually want to review a built-in default:

build/dist.js
generated/**/*.ts

# Re-include this lockfile when you want to review it
!package-lock.json

If the post-exclude diff is still over GIT_AI_MAX_DIFF_BYTES (default 900000, set 0 to disable), git-ai aborts with a "Largest changed files" hint pointing at what to ignore or unstage.

In the Python library, get_staged_diff, get_diff, and get_diff_stat auto-load .git-ai-ignore and apply built-in lockfile defaults when exclude_patterns is omitted. Pass exclude_patterns=[] to opt out of all filtering.

Repo-specific conventions (.git-ai-instructions)

git-ai's commit-type and scope heuristics are tuned for typical app repos. Some repos break those assumptions — a GitOps/deploy repo where every change is "user-facing" so the default feat-bias misfires, or a repo with its own commit scopes. Drop a free-form .git-ai-instructions file at the repo root to teach git-ai the local rules. Its contents are injected verbatim into the commit and PR prompts inside a <repo_guidance> block, and the prompts treat it as authoritative — when it conflicts with the built-in type heuristics, your guidance wins.

Lead with the repo's user POV. The single biggest lever on prefix accuracy is stating who consumes what this repo produces and what they perceive as its output. git-ai's prompts already reason from that POV — "user-facing" means visible to that audience — so the same edit can be feat in one repo and chore in another. Start with a User POV: line, then a few type rules expressed in that POV's terms. The examples below are illustrative — write your own repo's POV:

# .git-ai-instructions
User POV: the running app a deploy serves. "User-facing" = what changes for someone using it.

- Bumping an image tag to ship the same app's next build → chore
- A brand-new deployed service, or a capability its users gain → feat

# .git-ai-instructions
User POV: the report this tool prints. Its wording and layout are the product's UI.

- Rewording or restyling the printed report → style (or fix when correcting wrong output)
- A new export path that consumers of the report never see → build or ci, not feat

An absent or empty file is a no-op. In the Python library, load_repo_instructions(repo_path) returns the trimmed text (or None), and build_commit_prompt / build_mr_prompt accept a repo_guidance= argument.

Manual configuration (advanced)

Everything below is handled for you by git-ai setup. Reach for it only when you want to script config, edit it by hand, or set up an advanced case the wizard leaves alone (service accounts, multi-project Vertex profiles).

API keys by hand

For gemini-api, anthropic-api, and openai-api, git-ai resolves each key in this order until one succeeds:

1. The environment variable — GEMINI_API_KEY, ANTHROPIC_API_KEY, or OPENAI_API_KEY.
2. System keychain, under the service name <provider>-api-key (e.g. gemini-api-key, anthropic-api-key, openai-api-key):
   • macOS: security add-generic-password -s gemini-api-key -a "$USER" -w YOUR_KEY
   • GNOME / libsecret: secret-tool store --label="Gemini API Key" service gemini-api-key
   • pass: pass insert gemini-api-key
   • KDE Wallet: kwallet-query kdewallet -w gemini-api-key

For Google ADC / service-account credentials, use a vertex-gemini or vertex-anthropic method (gcloud auth application-default login or GOOGLE_APPLICATION_CREDENTIALS).

Narrowing the picker list

By default git-ai options enumerates every supported provider/model combo. Most users only have access to a couple. To restrict the picker to just the providers and models you actually use, drop a config file at $XDG_CONFIG_HOME/git-ai/options.conf (usually ~/.config/git-ai/options.conf):

[claude-code]
claude-haiku-4-5-20251001
claude-sonnet-4-6

[codex]
gpt-5.4-mini

# Empty sections hide these providers entirely
[vertex-gemini]

[vertex-anthropic]

• [provider] headers must be one of: vertex-gemini, vertex-anthropic, gemini-api, claude-code, anthropic-api, codex, openai-api. Unknown headers are silently dropped.
• Model IDs under a header are passed through to the provider verbatim, so you can list future model IDs (e.g. a newly released claude-sonnet-5-0) without waiting for a git-ai release.
• Delete the file to restore the full shipped catalog.
• See examples/options.conf for a starter.

Pinning a GCP account (Vertex)

git-ai setup writes project / region / account for a single Vertex provider for you. This section is the manual reference for that, plus the advanced cases the wizard leaves alone: service-account credentials=, the shared [vertex] block, and multi-project profiles.

A [vertex-*] section accepts optional key = value lines alongside its model IDs to pin which account and project that provider uses (these keys are not models and never appear in the picker):

[vertex-anthropic]
project     = acme-prod        # overrides $GOOGLE_CLOUD_PROJECT / $GOOGLE_VERTEX_PROJECT
region      = us-east5         # overrides $VERTEX_LOCATION (default us-central1)
account     = [email protected]      # token via `gcloud auth print-access-token --account=…`
credentials = ~/keys/sa.json   # or: point ADC at a service-account JSON (~ expands to $HOME)
claude-sonnet-4-6

• account= selects a human Google login (authenticate each once with gcloud auth login); credentials= selects a service-account JSON. Set one or the other — with neither, plain gcloud ADC is used.
• Config values override the corresponding environment variables.
• Before each call, git-ai prints the account/project it used to stderr (e.g. git-ai: Vertex account [email protected] · project acme-prod (us-east5)).

To choose between multiple projects/accounts from the picker, give each a profile suffix — [vertex-anthropic@<profile>]. Every profile becomes its own picker entry (labelled Vertex AI [<profile>]), so the same account across two projects is fully supported:

[vertex-anthropic@acme]
project = acme-prod
account = [email protected]
claude-sonnet-4-6

[vertex-anthropic@sandbox]
project = acme-sandbox
account = [email protected]
claude-sonnet-4-6

Pass one explicitly with git-ai commit vertex-anthropic@sandbox or git-ai pr vertex-anthropic@acme:claude-sonnet-4-6.

Terminal picker

Running git-ai commit or git-ai pr without a provider argument launches an inline fzf picker over the same provider/model combos Lazygit uses. History entries float to the top. Pass provider or provider:model to skip the picker. Flags still parse, so git-ai pr --base staging opens the picker then runs against the chosen base.

Set GIT_AI_NO_FZF=1 (or pipe stdout) to disable the picker for scripting. If fzf isn't installed, the tools fall back to the last saved choice.

Lazygit integration

Requires fzf on your PATH. Add the following under customCommands: in ~/.config/lazygit/config.yml:

customCommands:
  - key: "<c-g>"
    description: "AI commit message (git-ai + fzf)"
    context: "files"
    command: |
      choice=$(git-ai options commit | fzf --delimiter='|' --with-nth=2 --no-sort --tiebreak=index --prompt='git-ai> ') || exit 0
      git commit -m "$(git-ai commit "${choice%%|*}")" --edit
    output: terminal

Pressing <c-g> in the files panel opens an fzf picker showing every auth+model combo (plus reuse saved message when available). Typeahead narrows instantly; Enter commits with the generated message. Selections float to the top of the list on subsequent invocations.

Compatibility

git-ai does not depend on a specific terminal UI. It works in the CLI, in Lazygit, and in similar git environments as long as Git exposes the required repository state:

• git-ai commit needs staged changes (git diff --staged)
• git-ai pr needs commits and diff data relative to a base branch