One static binary. No Node. No Python. No Docker. No cloud account.
Point it at any OpenAI-compatible endpoint and you have a coding partner that reads and edits your code, runs your shell, verifies its own work, and remembers how you like things.
-000?style=flat-square&logo=apple&logoColor=white){kind=icon}
English · Tiếng Việt · 简体中文
liveaizen.mp4
# Windows (PowerShell)
irm https://raw.githubusercontent.com/talmetis-labs/aizen/main/install.ps1 | iex# Linux / macOS
curl -fsSL https://raw.githubusercontent.com/talmetis-labs/aizen/main/install.sh | shThen open a new terminal:
aizen account login # on an Aizen plan? sign in — no key to fetch, nothing to paste
aizen config # or bring your own: base URL → API key → pick a model
aizen # land in the REPL and start typingThat's the whole setup. No env vars, no config file to hand-edit.
An Aizen plan is bought by signing in: the session opens the gateway on its own, and nothing but the
token is written to disk. On a machine that cannot open a browser — SSH, a container, CI — pair it
instead with aizen login, which approves a short code and comes back with a key.
Prefer to do it by hand? Grab a binary from the
latest release — or build it yourself with
cargo install --git https://github.com/talmetis-labs/aizen. Upgrade or roll back any time with
aizen update. The Windows .exe is unsigned, so SmartScreen will ask: More info → Run anyway.
| One file, no runtime | 34 MB static binary, ~10 ms cold start. No Node, no Python, no 2 GB virtualenv. Runs on a 512 MB VPS, a scratch container, a CI runner, or a Pi. |
| Bring your own model | Any OpenAI-style /chat/completions endpoint — OpenAI, OpenRouter, a local llama.cpp/vLLM, an Anthropic gateway. Never locked to one lab. |
| It finishes the job | Reads, edits, runs your shell — then verifies before claiming done: it runs your typecheck and tests, and fixes what it broke. |
| It remembers you | An offline, BM25-ranked memory brain that learns from reuse — plus a persona, a durable SOUL identity, and skills it writes for itself after real work. |
| It runs where you aren't | aizen serve drives the agent from Telegram or Discord and asks your phone to approve risky edits. Host it on systemd, Docker, or Kubernetes — behind NAT, no inbound port. |
| Safe by construction | An OS sandbox under the approval layer: children never inherit your API keys, network is deny-by-default, and the filesystem policy is kernel-enforced on Linux (Landlock+seccomp) and macOS (Seatbelt) — Windows gets Job-Object containment and honest partial reporting, not pretend enforcement. A hard command floor refuses catastrophic commands even under auto-approve. aizen sandbox status shows exactly what your machine enforces — see docs/SANDBOX.md. |
aizen agent "fix the failing parse test"
⚙ search_files "fn parse_config" 3 hits
⚙ file_read src/config.rs 142 lines
⚙ file_edit src/config.rs 3 edits
⚙ shell_run cargo test ✓ 0 failed · 1.18s
verify gate passed
| Unified REPL | One chat + agent loop, no mode switch. Live HUD: model · tokens · turn · % context. Markdown, tables, diagrams, image input. |
| Agent loop | Parallel reads, approval-gated writes, LSP-powered symbolic edits, sub-agent dispatch, and a verify gate that must pass before "done". |
| Multi-agent | The Pantheon: seven capability-scoped sub-agents (argus finds · metis plans · daedalus builds · nemesis reviews · themis tests · clio researches · mnemosyne recalls); aizen workflow fans them out and synthesises one answer. |
| Web + browser | Search, fetch, and a katana-style crawler — all SSRF-guarded. Opt-in CDP tools drive a real Chrome. |
| Extensible | MCP servers (stdio/HTTP, OAuth 2.1), markdown slash-command macros, outbound notify channels. |
| Recoverable | Git-backed checkpoints — /timemachine rewinds a bad turn. |
→ Full reference — every command, the REPL surface, self-hosting, MCP, browser tools, and the safety model in detail.
Issues and PRs are welcome. Contributors agree to the CLA once — the project stays Apache-2.0, and §3 also grants the maintainer the right to license contributions commercially. Read it before you sign. See CONTRIBUTING.md.
Apache License 2.0 — open source, commercial use allowed. Keep the license and copyright notices, state your changes, and pass along the NOTICE file. Includes an express patent grant (§3).
"Aizen" and the logo are trademarks of the Aizen authors; §6 grants no trademark rights, so a fork must not present itself as Aizen. Releases up to v0.5.5 shipped under PolyForm Noncommercial; everything after is Apache-2.0.