agent6

A coding agent that jails model commands and uses editable state machines for long-running tasks.

The model can write code and ask to run commands, but those commands go through a jail with restricted filesystem and network access. Long-running workflows can be written, reviewed, edited, resumed, and replayed as declarative state machines instead of being left to an open-ended agent loop.

Full documentation: agent6.dev

Features

• Jailed commands: every command the model asks to run goes through a jail that controls what it can read and write and restricts its network access; auto picks the strongest level the host allows (Security)
• State machines: long-running workflows as declarative .asm.toml files you review, edit, test offline, run, watch, and replay, with waits, operator input, and steering built in (State machines)
• Three session kinds: run edits; plan and ask never do; --from <id> seeds one from another, and /btw asks a question beside a live run
• Verify gate: the repo's test command (inferred when unset) certifies the tree before a run may finish, and every surface shows the same green or red
• Clean checkout: every step commits to the run's own hidden ref, so your branch, HEAD, and index are never touched (a visible agent6/<id> branch tracks it by default); sessions merge lands it, resume continues it, fork branches it at any turn
• Task graph: the worker's plan is a persistent DAG, live on every surface and surviving crashes and compaction
• Context control: compaction is visible everywhere, /compact [focus] and /pin steer it, and repo memory carries lessons across runs
• Four front-ends, one engine: CLI, TUI, browser (desktop or phone), and editor over ACP all drive the same runs
• Live runs are addressable: attach follows and answers one, steer queues an instruction from a script or cron job, exec and forward reach inside its sandbox network, ps and history find it
• Parallel fan-out: --parallel N|model-a,model-b runs isolated lanes and compares them into a ranked report (Architecture)
• Code review: agent6 review on any diff, plus an in-loop adversarial panel where only blocking findings gate a finish
• Background commands: a run's command can keep running behind a handle (dev servers, watchers); none outlive the run
• Skills: SKILL.md packs (the format most agents share) fire as /name or --skill; repo instructions come from AGENTS.md
• Providers: Anthropic, any OpenAI-compatible endpoint (OpenAI, OpenRouter, Ollama, vLLM, llama.cpp, LM Studio), or a ChatGPT subscription, with model and reasoning effort set per role (Config)
• Budget: a hard max_usd cap per run, a token cap for calls with no price
• Fixed tool surface: the model's tools are a fixed set, extended only by operator-configured MCP servers (off by default, jailed by default)
• Eight runtime dependencies, no telemetry, no auto-update

Install

From PyPI with uv or pipx:

uv tool install agent6        # or: pipx install agent6

If agent6 is not found, you can add the uv or pipx bin dir (~/.local/bin) to your PATH with uv tool update-shell or pipx ensurepath.

Enable shell completion with agent6 completions (supports bash, zsh, fish, and xonsh).

agent6 requires Python 3.12+ and the sandbox only supports Linux (x86_64/aarch64). Other platforms run without the sandbox behind a warning. See installation for the full requirements and building from source.

Usage

# Connect a provider (stored in ~/.config/agent6/, key in a 0600 secrets file).
# If already connected, skip both; `agent6 check` verifies it.
agent6 connect                # interactive: pick provider, paste API key
# (or `agent6 connect chatgpt` to sign in with a ChatGPT subscription)
agent6 model worker anthropic claude-sonnet-5

# Run the agent on a task, create a plan, or ask a question.
cd your-repo
agent6 run "add a --json output mode to the CLI"
agent6 plan "how to add a --json output mode to the CLI"
agent6 ask "how to add a --json output mode to the CLI"

# Watch and drive runs from a terminal, a TUI, a browser, or an editor.
agent6 attach <session-id>    # follow + answer a run live (--raw for events)
agent6 tui                    # full-screen dashboard hub
agent6 web                    # browser UI on http://127.0.0.1:7658
agent6 acp                    # speak ACP on stdio; an editor spawns this

# Audit the effective config, check the sandbox, resume or fork a run.
agent6 config show
agent6 check
agent6 resume <session-id>
agent6 fork <session-id> --at-turn 7

# See all commands with `agent6 --help` or `agent6 <command> --help`.

See usage for the full command tour, the web UI for driving runs from a phone, configuration for every field, and the security model for what the sandbox enforces.

The general rules, which the rest of agent6 follows:

• Commands need your approval: under the default run_commands = "ask", each command the model proposes waits for your yes, once or for the session; a headless run auto-denies, and a hub-spawned one parks the prompt for a front-end
• A run never touches your checkout: its work lands as per-step commits on its own chain (a visible agent6/<id> branch by default), and agent6 sessions merge lands them when you are ready
• Config is layered, lowest precedence first: built-in defaults, the global ~/.config/agent6/config.toml, the per-repo config (state dir, never committed), then --config FILE
• Nothing is hidden: agent6 config show prints every effective value and the layer that set it; every field has a default, and security-sensitive fields default safe (network = "auto", run_commands = "ask", protect_git = true)
• "auto" picks the most secure option the host allows, warning when it falls short; an explicit value the host cannot enforce refuses to run
• agent6 never pushes, rewrites history, or reset --hard; no config key can enable them