Yes. Hazzel is free and open source (AGPL-3.0). You pay your model provider directly for API usage — or nothing at all with local Ollama models. No subscription.

Which AI models does Hazzel support?

Groq, OpenAI, Anthropic, Mistral, Gemini, DeepSeek, OpenRouter and local Ollama — one API key each, switched with /model.

Does Hazzel change files without asking?

No. Every file edit shows a diff before it applies, shell commands ask first, and every write is checkpointed so /undo always restores prior bytes.

What do I need to install Hazzel?

Python 3.10+ on macOS, Linux or Windows. pip install hazzel, run hazzel, add a key with /model.

How is Hazzel different from Aider or Claude Code?

Hazzel is ~10k lines of Python you can trace end to end, bring-your-own-key with no subscription, and approval-first on every mutation with automatic undo checkpoints.

Hazzel: the open source terminal coding agent

A tiny AI coding assistant that lives in your terminal — approval-first, bring your own key.

hazzel v1.5.7 · free · AGPL-3.0 · macOS, Linux, Windows · ★ star if useful

Hazzel is a terminal coding agent and CLI harness: it reads code, edits files, runs commands, reviews diffs and reads the web — always with your approval first. Every step is visible, every mutation goes through you.

It is open source (~10k lines of Python you can trace end to end), model-agnostic (Groq, OpenAI, Anthropic, Mistral, Gemini, DeepSeek, OpenRouter, local Ollama), and free — no subscription. You pay your provider directly, or nothing at all on local models.

Hazzel demo: ask it to fix the failing test in test.py — read, approve the edit with y, tests pass — Watch Hazzel fix the bug in test.py: read, diff approved with y, tests green.

Install the terminal coding agent

$ pip install hazzel

Requires Python 3.10 or newer. Then:

export GROQ_API_KEY="..."   # or skip this and pick a provider inside with /model
cd your-project
hazzel

The whole loop is agent/core.py, every action is in tools/, the undo system is safety.py — and Hazzel stops before every write to show you what is about to happen.

Hazzel 1.5.7 · ~/hazzel

❯ Fix the failing test in tests/test_agent.py

read_file · edit_file (diff approved) · run_command pytest -q — passed

openai/gpt-oss-120b · 1.2k tokens · $0.004 session · build · @ tag file

Why Hazzel

Approval-first: file changes show a diff before they apply, shell commands ask first, destructive ops are checkpointed so /undo always works.
Small and inspectable: no deploys, no hidden behavior — a terminal agent you can see through. Verify the safety claims in ~200 lines: src/hazzel/safety.py, src/hazzel/tools/run_command.py.
Real cost visibility: provider-reported tokens priced into dollars and logged locally — /usage today|week|month, /budget warns but never blocks.
Goal-driven: /goal pins an objective and every turn steers toward it until done.
Plan and think modes: /plan on explores read-only and hands you a numbered plan; /think on reasons step-by-step on hard edits.
Skills and AGENTS.md: drop in a SKILL.md or project map and Hazzel follows it every turn.
Sessions that survive restarts: per-project history saves automatically, /session restore picks up where you left off.
Model and provider agnostic: Groq, OpenAI, Anthropic, Mistral, Gemini, DeepSeek, OpenRouter, and local Ollama through one key each.

Hazzel vs alternatives

Looking for an Aider alternative or a Claude Code alternative you can run in any terminal with your own key? How the options compare:

Open source terminal AI coding assistants compared
Tool	Open source	Bring your own key	Approval-first diffs	Built-in undo	Local models
Hazzel	Yes (AGPL-3.0)	Yes — 8 providers	Yes, every write	Yes (`/undo`)	Yes (Ollama)
Aider	Yes (Apache-2.0)	Yes	Yes	Git-based	Yes
Cloud agents	Mostly no	Subscription	Varies	Varies	No

Hazzel's bet: most agents ask you to trust a black box. Hazzel asks you to trust three specific, inspectable mechanisms instead — approved diffs, automatic checkpoints, project-root sandboxing.

Commands

Modes

/model — switch provider and model
/plan on|off — read-only exploration with a numbered plan
/think on|off — deeper reasoning when a task is hard
/goal [@objective] — pin an objective, /goal run executes it

Cost

/usage — tokens and dollars this session
/usage today|week|month — spend over time, --by-model per model
/budget — warn-only spend limits, never blocks

Git

/review [--staged] — read-only AI review: Summary, Risks, Missing, Before commit
/commit — auto-drafted message with diff preview
/status · /diff · /log — git without leaving the loop

Transcript

/summary — what the last turn did
/export [file] — save the transcript to markdown
/copy [code] — copy the last reply
/retry — re-run your last message
/jobs [id|kill id] — background shell jobs
/undo [n] — revert file changes
/session restore — pick up where you left off
/clear — reset the conversation

Project

/init [file] — generate an AGENTS.md project map
/skills [name] — list installed skills, preview one
/help — quick overview, /docs — the full usage guide
/logout — wipe saved keys, /exit — quit

Type @path to attach a file, !command to run shell directly (& suffix runs it in the background — /jobs polls and kills).

hazzel -p "prompt" answers once and exits — pipe stdin in, -y allows writes, --output-format json for scripts.

changelog · PyPI · GitHub source

Frequently asked questions

Is Hazzel free?: Yes. Hazzel is free and open source (AGPL-3.0). You pay your model provider directly for API usage — or nothing at all with local Ollama models. No subscription, nothing metered by Hazzel.
Which AI models does Hazzel support?: Groq, OpenAI, Anthropic, Mistral, Gemini, DeepSeek, OpenRouter and local Ollama — one API key each, switched any time with /model.
Does Hazzel change my files without asking?: No. Every file edit shows a diff before it applies, shell commands ask first, and every write is checkpointed so /undo always restores prior bytes. Commands are sandboxed to your project root.
What do I need to install Hazzel?: Python 3.10+ on macOS, Linux or Windows. pip install hazzel, run hazzel, add a key with /model.
How is Hazzel different from Aider or Claude Code?: Hazzel is ~10k lines of Python you can trace end to end, bring-your-own-key with no subscription, and approval-first on every mutation with automatic undo. See the comparison table.