git-chat

Self-hosted LLM chat bound to a git repository. Single binary (Go + Lit). Works as a git chat subcommand or standalone CLI. Chat with an LLM about your codebase -- the knowledge base automatically invalidates when source code changes.

Quick start

# Install
go install github.com/pders01/git-chat/cmd/git-chat@latest

# Run (auto-detects repo, opens browser)
cd ~/code/my-repo
git chat

# Explicit backend
git chat --llm-backend=anthropic --llm-api-key=$ANTHROPIC_API_KEY

CLI usage

git chat                         # local mode, auto-detect repo, auto-open browser
git chat serve --repo /path/to/repo1 --repo /path/to/repo2
git chat local ~/workspace       # scan directory for git repos (multi-repo workspace)
git chat local --no-scan .       # treat CWD as single repo (fail if not valid)
git chat local --max-repos=5 .   # scan directory, load at most 5 repos
git chat mcp                     # MCP server mode (stdio)
git chat add-key paul@laptop < ~/.ssh/id_ed25519.pub

Multi-repo workspace support

When running git chat on a directory that contains multiple git repositories:

• Auto-scan: If the path is not a valid git repo, subdirectories are scanned for .git folders
• Explicit repos: Use --repo (repeatable) to specify exact repo paths, bypassing scan
• Limit loading: --max-repos=N caps the number of repos loaded from a scan
• Single repo mode: --no-scan treats the path as a single repo (fails if invalid)
• UI switcher: When multiple repos are loaded, a dropdown selector appears in the header, and Cmd/Ctrl+K command palette shows "Switch to: {repo}" actions

Features

• Chat -- streaming LLM responses with markdown rendering, @-mention file injection (recursive autocomplete), [[diff]] marker expansion, KB cache hits
• Composer slash commands -- /diff to hand-author [[diff]] markers, /model <id> and /profile <name> to switch LLM for the current chat, /help to list commands; arg autocomplete suggests branches/tags/paths for /diff, discovered models for /model, saved profiles for /profile
• Persistent model indicator -- small status line above the composer shows the model and profile the next turn will hit; refreshes instantly after /model or /profile
• Knowledge base -- FTS5 fuzzy matching, N>=2 promotion threshold, blob-SHA provenance, git-aware invalidation, webhook notifications (Slack/Discord) with exponential-backoff retry
• KB management -- card list with detail view, provenance display, delete, hit counts, invalidation status
• File browser -- expandable file tree (VS Code-style, lazy-load, arrow-key nav) + syntax highlighting (Shiki, 25 languages) + focus mode
• Git blame -- 2-panel info pane with interactive tooltip (cursor-following, 300ms debounce), "view in log" / "ask in chat" action buttons
• Commit log -- 3-column layout (list | info pane | diff pane), per-file diffs, SVG timeline graph via parent SHA linking
• Branch comparison -- file-by-file diff between any two refs
• Working tree changes -- staged, unstaged, untracked files with per-file diff viewer
• 3D code city -- Three.js visualization of file churn (commit frequency, additions/deletions, file size mapped to building dimensions), squarified treemap layout, time slider; hover tooltip surfaces filename/directory split plus the file's top author, keyboard navigable (arrows rotate, +/- zoom, Tab cycles buildings, Esc closes detail)
• File history -- commits filtered to a single file path
• Side-by-side diff -- LCS-based word-level diff highlighting via Shiki diff grammar
• Commit search -- filter by message or author
• Branch/tag switching -- global ref selector in header (branches + tags)
• Session pinning -- pinned sessions float to the top of the sidebar; star/unstar to toggle
• Streaming token counter -- live token count + cost estimate per turn
• LLM dashboard summaries -- auto-generated activity summary with suggested questions, rendered as markdown so list-shaped fallbacks become real bullets
• Runtime-configurable settings -- sidebar-navigated settings UI with LLM provider/model switching, effective-model status card (active profile / config override / compiled default), 3-tier resolution (SQLite override -> env var -> compiled default), API keys encrypted at rest (AES-256-GCM)
• Ad-hoc model discovery -- typing a custom base URL in the advanced config kicks a debounced /v1/models probe; results feed the model combobox so providers outside the catalog (Fireworks, Groq, custom deployments) just work
• Cross-tab navigation -- "explain in chat" from log, "ask in chat" from browser, blame-to-log SHA linking, Cmd+click files
• MCP server mode -- 10 tools: search_knowledge, get_file, get_diff, list_commits, search_files, search_code, outline, list_tree, list_branches, get_blame
• Two LLM backends -- OpenAI-compatible (LM Studio, Ollama, vLLM, OpenAI, Groq, Fireworks) + Anthropic native, switchable at runtime via settings UI
• Theme support -- system/light/dark with Shiki dual-theme
• Auth -- local mode (token URL) or SSH key pairing for multi-user

Architecture

Go backend (Connect-RPC, connectrpc.com/connect) + Lit frontend (embedded via go:embed). Single .proto source of truth generates both Go server handlers and TypeScript clients. All state in one SQLite file (modernc.org/sqlite, pure Go, no CGO).

See docs/ARCHITECTURE.md for the full design document.

Environment variables

LLM

Variable	Default	Description
LLM_BACKEND	openai	openai or anthropic
LLM_BASE_URL	http://localhost:1234/v1	OpenAI-compatible endpoint (ignored for anthropic)
LLM_MODEL	(empty)	Model name; empty falls back to the backend default
LLM_API_KEY	--	API key (required for anthropic)
LLM_TEMPERATURE	(empty)	Sampling temperature; empty = use startup flag, 0 = deterministic
LLM_MAX_TOKENS	(empty)	Max tokens per response; empty = use startup flag, 0 = provider default
LLM_ACTIVE_PROFILE	(empty)	Active LLM profile ID; empty = use individual LLM_* settings
LLM_SYSTEM_PROMPT	(empty)	Custom system prompt snippet (appended to base prompt)

Repo

Variable	Default	Description
GITCHAT_MAX_DIFF_BYTES	524288	Max whole-commit diff size
GITCHAT_DEFAULT_COMMIT_LIMIT	50	Default commit list page size
GITCHAT_DIFF_CONTEXT_LINES	3	Context lines around diff hunks
GITCHAT_DEFAULT_FILE_BYTES	524288	Default max file size for GetFile

Chat / Prompt

Variable	Default	Description
GITCHAT_MAX_FILE_BYTES	4096	Per @-file injection cap
GITCHAT_MAX_TOTAL_INJECT	12288	Total @-file budget per turn
GITCHAT_MAX_BASELINE_BYTES	4096	Overview doc injection cap
GITCHAT_MAX_HISTORY_DIFF	4096	Per-diff cap in history expansion
GITCHAT_MAX_TREE_LINES	60	Max lines in baseline tree listing
GITCHAT_MAX_TREE_BYTES	2048	Max bytes in baseline tree listing
GITCHAT_RECENT_COMMITS	5	Recent commits in system prompt
GITCHAT_MAX_MESSAGE_BYTES	32768	Max user message size
GITCHAT_MAX_HISTORY_TURNS	20	Sliding window of history turns sent to LLM
GITCHAT_TITLE_MAX_LEN	48	Max auto-generated session title length
GITCHAT_TITLE_TIMEOUT	15s	Timeout for LLM title generation
GITCHAT_CARD_TIMEOUT	10s	Timeout for KB card promotion
GITCHAT_KB_PROMOTION_THRESHOLD	2	Min similar questions before promotion

Session

Variable	Default	Description
GITCHAT_SESSION_TTL	168h	Browser session cookie lifetime
GITCHAT_DEFAULT_SESSION_LIMIT	100	Default session list page size

Webhook

Variable	Default	Description
GITCHAT_WEBHOOK_URL	(empty)	Slack/Discord incoming webhook URL (empty = disabled). Sends card_invalidated and card_created events; transient failures (network, 5xx, 429) retried with exponential backoff up to 3 attempts.

Storage

Variable	Default	Description
GITCHAT_DB	~/.local/state/git-chat/state.db	SQLite state file
XDG_STATE_HOME	~/.local/state	Base for default DB path

All GITCHAT_* and LLM_* variables are runtime-configurable via the settings UI. Values set in the UI are stored as SQLite overrides and take precedence over env vars. Secret values (LLM_API_KEY) are encrypted at rest with AES-256-GCM. Sensitive keys (LLM_API_KEY, LLM_BASE_URL, GITCHAT_WEBHOOK_URL) can only be changed in local mode.

Keyboard shortcuts

Key	Action
Cmd+K / Ctrl+K	Command palette
Cmd+F / Ctrl+F	Search
Double-press Cmd+F / Cmd+K	Fall through to the browser native (find / URL bar) — press twice within 300ms
Cmd+1-4 / Ctrl+1-4	Switch tabs (chat, browse, log, kb)
Cmd+\ / Ctrl+\	Toggle focus mode
/	Focus composer (or open the slash-command menu inside the composer)
Esc	Blur / close modal
?	Shortcut help
F2	Rename selected session
ArrowUp/Down	Navigate lists (sessions, files, commits, mentions, slash menus)
Tab	Accept suggestion and keep editing (slash arg menus)
Enter	Select / expand / accept + submit
ArrowRight/Left	Expand / collapse file tree nodes
Code-city Arrows / + - / Tab / Shift+Tab / Esc	Orbit camera / zoom / cycle buildings / close detail

Development

make dev          # vite HMR on :5173 + Go server on :8080
make check        # go vet + go test + tsc + oxlint + oxfmt
make test-e2e     # Playwright (desktop + mobile), 29 tests
make size-check   # fails if embedded SPA exceeds MAX_BUNDLE_BYTES (default 6 MiB)
make proto        # buf generate
make all          # frontend + Go binary

Frontend unit tests (via bun:test, happy-dom harness): cd web && bun run test.

Requirements: Go 1.25+, Bun 1.2+, Chromium (for Playwright).