What is Autonomous Company?

Autonomous Company is an MIT-licensed framework that runs a stand-up-your-own-LLM-team loop on a single machine: a daemon keeps calling a coding-agent CLI in cycles, and each cycle reads a shared consensus memory file, decides what to do next, assembles a team from fourteen expert-persona role definitions (CEO, CTO, product design, full-stack, QA, marketing, CFO and more), executes research, coding or marketing work, and writes the summary back before sleeping. Work is steered by editing the consensus file between runs, and everything stays inside human-configured goals, permissions and cost budgets, with per-cycle token and USD ledgers that can hard-pause the loop.

A local web Product Center on loopback port 8810 gives a catalog of connected product runs, read-only import of existing run directories, a work queue that runs up to four projects in parallel by default, batch creation with reusable engine/model/effort templates, and visible failure and recovery states. The loop itself ships resilience: circuit breakers on consecutive errors, backoff on rate limits, consensus rollback, and per-cycle process supervision with timeouts. Engine support is Claude Code by default and Codex CLI, with opt-in Cursor and OpenAI-compatible adapters; execution requires a Linux namespace boundary built on bubblewrap, so model cycles run under WSL2 on Windows and native Windows execution is refused.

Setup is clone the repository and use make targets (start, install as a launchd or systemd --user daemon, status, monitor, pause, resume, dashboard, center) or the PowerShell entrypoints on Windows; you must install and authenticate a supported agent CLI and pay for its model quota. The project describes itself as experimental: it costs money per cycle, prompt-level guardrails are not an operating-system sandbox, and stability is not guaranteed. It ships 22 example generated applications and English plus Chinese documentation.

Should you self-host Autonomous Company?

Autonomous Company scores 59/100 on selfhost.directory (how scoring works) and ranks #54 of 209 in AI Agents. It has 3 160 GitHub stars.

It shipped 10 stable releases in the past year. The latest, v3.0.0, came out on 13 October 2026. New versions land roughly every 1 days.

Setup is advanced: plan for several services and some ongoing maintenance. It is MIT-licensed and written mostly in Python.

Its closest alternative on selfhost.directory is Orca (91/100, 108 446 stars). It scores higher overall, so compare the two before you pick. See Autonomous Company vs Orca side by side.

Features

14 expert agent role personas
Background daemon with auto-restart
Batch work creation with templates
Circuit breaker and rate-limit backoff
Continuous autonomous work loop
English and Chinese interface
Human-in-the-loop steering
Local Web Dashboard
Multi-product center with read-only import
Parallel project executionup to four projects at once by default
Pluggable agent engine adapters
Reusable skill library30+ skills
Sandboxed project execution
Shared file-based memory across cycles
Usage and cost budget governance

How to configure Autonomous Company

AUTONOMOUS_COMPANY_BROWSER_RUNTIMENames a reviewed Playwright browser directory mounted read-only into the isolated execution.
Names a reviewed Playwright browser directory mounted read-only into the isolated execution.
AUTONOMOUS_COMPANY_BWRAPPoints the preflight at an operator-installed bubblewrap executable.
Points the preflight at an operator-installed bubblewrap executable.
AUTONOMOUS_COMPANY_LANGUAGESupplies the initial loop language when no saved preference exists.
Supplies the initial loop language when no saved preference exists.
values: en | zh-CN
AUTONOMOUS_COMPANY_LIBRARY_RUNTIMEReviewed directory of shared libraries needed by Chromium inside the sandbox.
Reviewed directory of shared libraries needed by Chromium inside the sandbox.
CLAUDE_BINOptional override path to the Claude Code binary.
Optional override path to the Claude Code binary.
CLAUDE_PERMISSION_MODEPermission mode passed to the Claude Code CLI.
Permission mode passed to the Claude Code CLI.
values: acceptEdits, auto, bypassPermissions, default, dontAsk, plan
CODEX_BINOptional override path to the Codex CLI binary.
Optional override path to the Codex CLI binary.
CODEX_REASONING_EFFORTReasoning effort passed to the engine (shown with the Codex engine).
Reasoning effort passed to the engine (shown with the Codex engine).
CODEX_SANDBOX_MODEOptional Codex inner sandbox policy override.
Optional Codex inner sandbox policy override.
CURSOR_ADAPTER_ENABLEDExplicit opt-in switch for the Cursor engine adapter.
Explicit opt-in switch for the Cursor engine adapter.
values: 1
CURSOR_BINPath to the cursor-agent binary for the Cursor adapter.
Path to the cursor-agent binary for the Cursor adapter.
CYCLE_TIMEOUT_SECONDS1800Timeout for a single cycle in seconds.
Timeout for a single cycle in seconds.
ENGINEclaudeExecution engine selection for the loop.
Execution engine selection for the loop.
values: claude | codex | cursor | openai-compatible
LOOP_INTERVAL30Seconds to wait between loop cycles.
Seconds to wait between loop cycles.
MAX_CONSECUTIVE_ERRORS5Circuit-breaker threshold of consecutive errors.
Circuit-breaker threshold of consecutive errors.
MODELOptional model override for the loop.
Optional model override for the loop.
OPENAI_COMPATIBLE_ADAPTER_ENABLEDExplicit opt-in switch for the OpenAI-compatible engine adapter.
Explicit opt-in switch for the OpenAI-compatible engine adapter.
values: 1
OPENAI_COMPATIBLE_ALLOW_INSECURE_HTTPExplicitly allows plaintext non-loopback HTTP for the OpenAI-compatible endpoint; disabled by default.
Explicitly allows plaintext non-loopback HTTP for the OpenAI-compatible endpoint; disabled by default.
values: 1
OPENAI_COMPATIBLE_ALLOW_SHELLGives the OpenAI-compatible model an arbitrary Bash tool, bypassing the file-tool path policy.
Gives the OpenAI-compatible model an arbitrary Bash tool, bypassing the file-tool path policy.
values: 1
OPENAI_COMPATIBLE_API_KEYAPI key for the OpenAI-compatible endpoint; read only from the process environment and sent as a bearer token.
API key for the OpenAI-compatible endpoint; read only from the process environment and sent as a bearer token.
OPENAI_COMPATIBLE_ENDPOINTExact chat-completions endpoint used by the OpenAI-compatible adapter; no built-in value.
Exact chat-completions endpoint used by the OpenAI-compatible adapter; no built-in value.
OPENAI_COMPATIBLE_MODELModel name used by the OpenAI-compatible adapter (MODEL also accepted).
Model name used by the OpenAI-compatible adapter (MODEL also accepted).
OPENAI_COMPATIBLE_SAFE_COMMANDSComma-separated allowlist of exact read-only commands the OpenAI-compatible adapter may run.
Comma-separated allowlist of exact read-only commands the OpenAI-compatible adapter may run.
USAGE_BUDGET_PERIODWindow over which usage budget thresholds are counted.
Window over which usage budget thresholds are counted.
USAGE_HARD_LIMIT_TOKENSHard token limit that pauses the next cycle.
Hard token limit that pauses the next cycle.
USAGE_HARD_LIMIT_USDHard limit in USD; the current cycle finishes, then the next cycle is blocked.
Hard limit in USD; the current cycle finishes, then the next cycle is blocked.
USAGE_WARNING_TOKENSWarning threshold in tokens.
Warning threshold in tokens.
USAGE_WARNING_USDWarning threshold in USD; only emits a BUDGET alert.
Warning threshold in USD; only emits a BUDGET alert.