Show HN: Hazzel – A tiny, minimal, Git-native coding agent

Hazzel 0.1.2, a minimal Git-native coding agent that runs in the terminal, is now available via pip install hazzel. It offers ten tools for code inspection, editing, and Git operations, supports bring-your-own API keys from Groq, OpenAI, Mistral, or Anthropic, and requires Python 3.10+. The agent emphasizes transparency, with every read, edit, and command visible, and includes optional plan and prove modes.

A small terminal coding agent that works on your code, not around it. Once upon a time, every coding assistant wanted to become your IDE, your cloud, your everything. Hazzel chose a different story. It stays where the work happens — your terminal, your project, your rules — and talks to your code through a small set of honest tools. You ask in plain language. It inspects, edits, runs, verifies, and shows its work. Nothing hidden. Nothing magic. Just you, your terminal, and an agent that knows its place. Hazzel 0.1.2 ~/hazzel ❯ fix the bug in agent.py ● read file agent.py ● edit file agent.py ● run command python -m pytest — passed Hazzel Fixed. Verified with tests. Most agents grow into platforms. Hazzel shrinks into a tool. The philosophy: 1. Small beats complicated. If it can be clear without another abstraction, keep it simple. 2. The model requests, Hazzel decides. The model never touches your machine directly. 3. The terminal is enough. No GUI, no cloud workspace, no lock-in. 4. Your keys are yours. Bring your own Groq / OpenAI / Mistral / Anthropic key. 5. Show the work. Every read, edit, and command is visible. pip install hazzel Requirements: Python 3.10+, a terminal, and one API key. From source: git clone https://github.com/mukundzha/hazzel.git cd hazzel pip install -e . Start Hazzel from the project you want it to work on: cd my-project hazzel python -m hazzel works too. That directory becomes the project root. File tools cannot escape it — ../another-project is rejected. No export needed. Just run /model , pick a model, and paste your key in the input field when asked — it's saved securely and is the most stable way to stay signed in. Config lives at ~/.config/hazzel/config.json 0600 . Ten tools — six for code, four git-native: | Tool | Story | |---|---| | list files "." | Look around the room | | read file "src/main.py" | Read 60 numbered lines, page with offset | | search files "MAX ITERATIONS" | Grep first, read second — never wander blind | | write file "src/example.py", "..." | Create, with undo | | edit file "src/main.py", "old", "new" | Small anchored edit, diff preview, undoable | | run command "python -m pytest" | Run shell, with approval, timeout, and cap | | git status | Branch + working-tree status, read-only | | git diff | Per-file diff with +/− counts, read-only | | git commit "msg" | Diff preview + approval; omit msg to auto-draft from diff | | git branch "list" | List, create, switch — writes ask first | Tag files directly: @src/agent.py fix this . Tab-completes, highlights, attaches to context. /model — switch model / provider /plan on — read-only plan mode: Hazzel explores and proposes, you approve off by default /prove on — ephemeral smoke check after edits off by default /status — git working-tree status /diff — changed files, open any file's full diff /commit — suggest a message from your diff, y/e/n /branch — list / create / switch branches /push — push branch to remote asks first /pull — pull remote changes asks first /sync — pull then push asks first /log — recent commits /help — shortcuts + commands /clear — reset conversation + usage /summary — what did Hazzel just do? /usage — tokens sent, received, cached /undo 2 — undo last 2 file changes /logout — clear saved keys /exit — leave Prove mode is optional and off initially. Turn it on with /prove on : after each edit Hazzel shows what it will run and asks once Run? , then either writes a happy-path smoke script to /tmp Python changes, imports just work or runs your repo's suite npm test , cargo test , go test , pytest for other files — and reports prove … — passed/failed . Nothing is written into your project. Plan mode is read-only and off initially. Turn it on with /plan on : Hazzel explores and answers with a numbered plan instead of acting — writes, runs, and commits stay blocked until you approve with /plan off . The footer shows · plan vs · build so the mode is always visible. The / menu pages 5 commands at a time with a n/total counter. Commands ask first: Hazzel wants to run: python -m pytest Allow? y/N Edits show a diff and ask first. Deletes and overwrites are checkpointed for /undo . Answers stream in token-by-token on all four providers, with automatic fallback to a full response if a stream drops mid-turn. The agent is frugal by design: trivial turns hi , read foo.py , list files run locally with zero LLM calls, history is compacted to ~2400 tokens, tool output is distilled to ~5000 chars. Pick your pilot with /model : - OpenAI: GPT-6 Astra, GPT-5.6 Sol/Terra/Luna, GPT-5, GPT-5 Mini, GPT-4.1, GPT-4o - Anthropic: Claude Fable 5.1, Opus 5, Sonnet 5, Haiku 4.5 - Mistral: Medium 3.5, Small 4, Large 3, Ministral 3, Codestral, GLM 5.2, Leanstral 1.5 - Groq: GPT OSS 120B default , GPT OSS 20B Default: openai/gpt-oss-120b on Groq. You ↓ main .py — listens, routes slash commands ↓ agent.py — fast-path + tool loop max 114 , history, budget ↓ providers/ — groq / openai / mistral / anthropic, one interface ↓ tools/ — the only hands that touch your project ↓ safety.py — checkpoint before every mutation ↓ ui.py + formatter.py — spinners, streaming answers, diffs, markdown, token meter The model decides what . Hazzel decides whether and how . Tools do the work. src/hazzel/ main .py agent.py config.py llm.py mentions.py safety.py tokens.py git.py git suggest.py ui.py formatter.py providers/base.py groq.py openai.py mistral.py anthropic.py tools/list files.py read file.py search files.py write file.py edit file.py run command.py git status.py git diff.py git commit.py git branch.py Hazzel is an agent, not a sandbox. run command runs on your machine — only approve what you understand. File writes/edits/ rm are undoable. Nothing leaves the project root without your y . 0.1.4 — early, deliberate, reliable foundation. Not big, on purpose. Included: terminal UI, multi-provider, tool-calling, @mentions, history compaction, undo, approval, timeout, token usage, markdown rendering, streaming answers, plan mode, git-native status/diff/commit/branch with message suggestions. Not yet: background commands, session persistence, IDE plugins. Omissions, not oversights. Built by Mukund Jha — mukundzha33@gmail.com mailto:mukundzha33@gmail.com License: AGPL-3.0-or-later — see LICENSE /mukundzha/hazzel/blob/main/LICENSE .