agent6¶
A sandboxed coding agent for Linux, tuned to stay effective even on affordable open-weight models like Kimi, GLM, and Qwen, as well as Claude.
Run agent6 with a weaker or untrusted model on any repository: it cannot escape the workspace, reach the network beyond your model provider, or corrupt git history. The model reads, searches, and edits files, runs the project's verify command, and commits each step that passes; state is snapshotted before every model call, so an interrupted or wrong run is resumable.
Per-command sandbox¶
Each command the model runs is jailed on its own (user namespaces, Landlock, seccomp,
pivot_root), with .git rebound read-only and egress limited to your provider. Not one
coarse container around everything.
Open-weight or frontier¶
Works with Anthropic and any OpenAI-compatible endpoint (OpenAI, OpenRouter, Ollama, vLLM, llama.cpp, LM Studio), and the prompts and tools are tuned to stay usable on cheaper open-weight models.
Resumable and forkable¶
State is snapshotted before every model call and committed after every passing step. Resume an interrupted run from its snapshot, or fork a new run from any past turn while the original stays intact.
Plan, run, review, ask¶
Read-only planning and repository Q&A, a diff-review panel, and the run loop, each its own command. A live terminal dashboard, full transcripts, and searchable run history.
State machines¶
agent6 machine composes longer automated tasks from runs, sandboxed tool calls, waits,
and branches: drafted by the model, reviewed by you, journaled, and replayable.
Small, fixed tool surface¶
The model's tools are a fixed set declared in one file. The only way to add more is an operator-configured MCP server, off by default. No telemetry, no auto-update.
The terminal UI¶
agent6 tui opens the hub: every run for the repository, with its mode, status, and
cost. Open a run to watch the dashboard, read the full transcript, or scroll the event
log. agent6 run opens the dashboard directly; --no-tui and -i (a stdin REPL) opt
out. The tour has a still of each screen.
Install¶
agent6 needs Linux for the sandbox (kernel 6.7 or newer for the network rules), Python 3.12 or newer, and an API key for at least one provider. macOS and Windows run unsandboxed behind a startup warning. See installation for the full requirements.
Run¶
agent6 connect # pick a provider, paste an API key (once)
agent6 model worker anthropic claude-sonnet-4-6
cd your-repo
agent6 run "add a --json output mode to the CLI"
agent6 infers a verify command when you have not set one, commits each step that passes it, and stops when the run finishes or a budget ceiling is hit. The getting started guide covers the first run and recovering one that went wrong.