agent6

A sandboxed coding agent for Linux. The LLM is treated as adversarial: every command it spawns runs inside a custom Rust launcher (agent6-jail) built on user namespaces, Landlock, seccomp, pivot_root, capset(0), and NO_NEW_PRIVS, so a misbehaving model cannot escape the workspace, reach the network beyond the provider endpoint, or corrupt git history.

Features:

• Sandboxed execution for every LLM-chosen child process (verify commands, metric commands, optional shell)
• Works with Anthropic and any OpenAI-compatible endpoint (OpenAI, OpenRouter, Ollama, vLLM, llama.cpp, LM Studio), tuned to stay effective on cheap open-weights models
• Per-step git commits, snapshot-resumable runs, USD and token budgets with hard stops
• Plan, run, review, and ask modes; a live terminal dashboard; persistent transcripts and a searchable run history
• State machines (agent6 machine) for long-running automated tasks: LLM-drafted, operator-reviewed, journaled, and replayable
• Small, fixed LLM tool surface; the only extension point is operator-configured MCP servers, off by default
• Eight runtime dependencies, no telemetry, no auto-update

Requirements

• Linux. The sandbox uses Linux-only kernel APIs; macOS and Windows are not supported.
• Kernel 6.7 or newer for Landlock TCP rules. Older kernels fall back to filesystem-only Landlock with a warning.
• kernel.unprivileged_userns_clone = 1 for the strict profile (default on Ubuntu, Debian, and most cloud images); without it agent6 falls back to hardened. On Ubuntu 24.04+ with kernel.apparmor_restrict_unprivileged_userns = 1, install the bundled AppArmor profile (packaging/apparmor/agent6-jail; agent6 check sandbox prints the commands) or set that sysctl to 0.
• Python 3.12 or newer, plus an API key for at least one provider.
• Building from source needs a Rust toolchain on PATH; PyPI wheels bundle a prebuilt agent6-jail.

Install

From PyPI with uv or pipx:

uv tool install agent6
pipx install agent6

Both drop the agent6 entry point in ~/.local/bin; if that is not on your PATH, run uv tool update-shell or pipx ensurepath and restart your shell.

From source:

git clone https://github.com/elesiuta/agent6
cd agent6
uv sync
uv run agent6 --help

AGENT6_JAIL_BIN=/path/to/agent6-jail overrides the bundled jail binary.

Shell tab-completion

Via argcomplete:

# Bash / Zsh
eval "$(register-python-argcomplete agent6)"

# Fish
register-python-argcomplete --shell fish agent6 > ~/.config/fish/completions/agent6.fish

Quick start

# Connect a provider once (stored in ~/.config/agent6/, key in a 0600
# secrets file). Works across every repo.
agent6 connect                # interactive: pick provider, paste API key
agent6 model worker anthropic claude-sonnet-4-5

# In a project: scaffold .agent6/config.toml + AGENTS.md.
agent6 init

# Audit the effective config: every value and where it came from.
agent6 config show

# Pre-flight: sandbox + config + provider keys + verify_command.
agent6 check

# Run the agent on a task.
agent6 run "add a --json output mode to the CLI"

# Resume an interrupted run from its last tool-call snapshot.
agent6 resume <run-id>

# Read-only code review of a diff. Never touches the worktree.
agent6 review --base origin/main --head HEAD

Config is layered: built-in secure defaults, then the global ~/.config/agent6/config.toml, then the per-repo .agent6/config.toml, then an explicit --config FILE. A repo can be zero-config when the global config supplies a provider and model; the one thing a repo always needs is its verify_command.

Other commands:

• agent6 watch [<run-id>]: attach the live TUI to an existing run.
• agent6 plan "<task>": read-only planning pass; execute with agent6 run --from-plan.
• agent6 ask "<question>": read-only Q&A over the repo, including questions about agent6 itself (it consults its bundled docs). Seed context with @path or --file; --run <id> asks about a prior run.
• agent6 memory: persistent agent memory under .agent6/memories/.
• agent6 history search <query>: search persisted transcripts.
• agent6 history graph [<run-id>]: render the persisted task graph.
• agent6 diff [<run-id>]: print the git diff a run produced.
• agent6 machine ...: author and run state machines (.asm.toml); see STATE_MACHINES.md.
• agent6 mcp serve: expose agent6's tools over MCP (stdio).
• agent6 config fill: materialize every effective value into one file.
• agent6 config get/set/unset/add/remove <key> [value]: edit a single dotted leaf. Writes go to the global config by default, --repo for the repo, --machine FILE for a machine overlay. Every edit is re-validated and rolled back if invalid.

Configuration

Every field has a default, and security-sensitive fields default to the safe value. The full reference is CONFIG.md; sandbox profiles are explained in SECURITY.md.

[sandbox]
profile = "auto"              # auto | strict | hardened
agent_network = "providers"   # providers | local | open  (agent's LLM egress)
tool_network = "block"        # block | only_explicit_states | allow  (jailed commands)
run_commands = "ask"          # yes | no | ask
protect_git = true
protect_agent6 = true

[git]
require_clean_worktree = true
branch_per_run = true
commit_strategy = "per_step"  # per_step | squash | stage | none
allow_push = false

[workflow]
verify_command = ["uv", "run", "pytest", "-x"]

[budget]
max_input_tokens  = 2000000
max_output_tokens = 200000
# best_effort_usd_limit = 10.0  # optional; see CONFIG.md

[providers.anthropic]
kind = "anthropic"
api_key_env = "ANTHROPIC_API_KEY"

[models.worker]
provider = "anthropic"
model = "claude-sonnet-4-5"

Budget ceilings can be overridden per run: agent6 run --max-usd 5 "...", or --max-input-tokens / --max-output-tokens on run, plan, and resume.

Providers and models

Declare any number of [providers.<name>] blocks, each with kind = "anthropic" or kind = "openai", its own base_url, and api_key_env. Per-provider http_timeout_s (default 600) caps each HTTP call.

agent6 uses three model roles:

Role	Routed by	Used by
worker	[models.worker]	agent6 run / resume; drives USD-to-token conversion.
reviewer	[models.reviewer]	agent6 review and the optional in-loop critic.
planner	[models.planner]	agent6 plan. Falls back to worker when unset.

agent6 model all <provider> <model> sets every role at once. Each role takes an optional thinking level (off/low/medium/high).

Tool surface

The tools given to the LLM are fixed and declared in one place, src/agent6/tools/schema.py; adding one requires a security review note in the commit message.

• Read-only: read_file, list_dir, grep, outline, find_definition, find_references
• Edits: apply_edit (structured blocks), apply_patch (unified diff)
• Execution with operator-fixed argv: run_verify_command, run_metric_command
• Control: finish_run, plus dag_* task-notepad tools backed by the curator subprocess
• Conditional: run_command(argv), only exposed when sandbox.run_commands allows it, and always jailed

There is no write_file, shell, or web_fetch.

How it works

agent6 is a single-loop agent: one provider, one model, one message history. The model drives the run by calling tools; the workflow dispatches them, snapshots state before every LLM call (so any run is resumable), commits when verify_command passes, and hard-stops on budget. Module boundaries (cli -> workflows -> agents -> tools -> sandbox) are enforced by tach. See ARCHITECTURE.md for the run/review loops, the curator subprocess, and the on-disk layout.

For security details (threat model, per-layer breakdown, sandbox profiles), see SECURITY.md. Defaults are safe: agent_network = "providers", tool_network = "block", run_commands = "ask", protect_* = true, git.allow_* = false, and git_ops.py refuses push, --force, and history rewrites unconditionally.

Benchmarks

Reproducible harnesses live under bench/:

• bench/realworld/: 11 SWE-bench-Lite-style tasks scored by fresh sandboxed verifies on hidden tests. Latest recorded run: agent6 and claude-code both solve 11/11 on the same worker model (claude-sonnet-4-5); agent6 at about $2.60 total, claude-code at about $3.96. Single runs, no variance measured; re-run before quoting.
• bench/agents/: head-to-head against Claude Code, opencode, and aider on Go and Rust tasks with cheap models.
• bench/machine/: machine create attempts, cost, and validation results.
• bench/perf/: a perf-optimization harness for local experimentation; single-run numbers are too noisy to quote.

Cost accounting

Every run ends with a per-model token and cost summary. Model prices are fetched from the provider's models endpoint and cached (OpenRouter publishes them; Anthropic does not, so its models report an unknown price). Where the provider reports per-call cost, that figure is used directly. The [budget] ceilings hard-stop the run; a stopped run is resumable.

Live view

With stdout a TTY, agent6 run opens a terminal dashboard (task DAG, budget bar, tool table, live reasoning pane, log tail, latest diff); --no-tui and -i (stdin REPL) opt out. Approval and Ctrl-C steer prompts appear as modals, with a /dev/tty fallback when no TUI is present. agent6 plan, agent6 ask, and agent6 machine create stream reasoning and answers to the terminal. Attach from another shell with agent6 watch [<run-id>]; agent6 watch --plain is a plain-text tail. The dashboard renders the JSONL event stream at .agent6/runs/<run-id>/logs.jsonl, which is also the contract for external viewers (vocabulary in ARCHITECTURE.md).

Persistence

Each run's state lives under .agent6/runs/<run-id>/: append-only task graph, per-call snapshots that drive agent6 resume, full transcripts, and the event log. It is written exclusively by a sandboxed agent6-curator subprocess over a validated IPC channel, so a bug in the agent cannot scribble the run directory.

End-of-run notify hook

If [notify] declares on_complete = [...], agent6 runs that argv after every agent6 run / resume with AGENT6_RUN_ID, AGENT6_RUN_DIR, AGENT6_RUN_OK, and AGENT6_RUN_REASON set. The hook runs outside the jail as your user; the argv is operator-controlled.

Contributing

Read AGENTS.md first. The repo's verify command decides whether a PR is landable:

uv run ruff check && uv run ruff format --check && \
  uv run pyright && uv run tach check && uv run pytest

Changes under sandbox/, tools/, git_ops.py, providers/, or graph/curator must include a security review note in the commit message.

License

Apache-2.0.