Autonomous AI Agency
An AI company that runs itself, on your servers.
Paste a website URL and get a CTO-grade audit in minutes. Then a CEO agent and a fleet of
specialists get to work: they plan, write code, test it, and open PRs, around the clock.
The only thing they ask you for is the merge.
Self-hosted · MIT · Your servers, your models, your data
▶ Try it live, free · ⚡ Run it in one command · 🤪 Meet SAM · 🧰 Agency Harness · 🔬 See the proof · 💬 Get help
▶ Watch the tour (61 s): paste a URL → stack scan → audit → specialist fleet → plan, execute, verify → your approval → 24×7 → meet SAM → it keeps learning → autonomy on a leash.
In 30 seconds
| |
|---|
| 🔎 It reads your business | Paste a URL. It scans the stack and runs a deterministic SEO / GEO / AIO / security audit with a 0–100 score and every finding priced. |
| 🏢 It staffs itself | It provisions the specialist agents that business needs, coordinated by a CEO agent with a C-suite advisory board (CFO, CSO, COO, CMO, CPO, Counsel). |
| 🛠 It ships, and checks its own work | Every change goes through plan → execute → verify. It byte-compiles and runs the tests before accepting a step. An LLM grading its own diff doesn't count. |
| 🧯 It fixes itself, and learns | Errors on the dashboard become fix tasks automatically. Failures turn into lessons the planner reads next time. Every hour it mines its own sessions for failures that keep repeating and files them as improvements. The CEO stops repeating directives that keep failing. |
| 🛡 It stays on a leash | Approval gates, a kill switch, per-agent daily spend caps, a canary credential, a policy engine and an audit trail. Nothing merges or deploys without you. |
| 🤪 It has a face | SAM, a googly-eyed little creature, floats in the corner of every screen. Tell it what to do and it runs the agency. |
🤪 Meet SAM
SAM (System Autonomy Manager) is the googly-eyed avatar in the corner of every screen and your
one-tap line to the agency. Type or speak to it: "brief me", "create a task to add dark mode to
billing", "triage the queue", "fix the alerts", or anything else in plain English. Commands that
create or approve work run as deterministic code, and loosening a safety control waits for your
Confirm tap. Realtime voice runs over LiveKit. SAM is on for
admins by default (Settings → Platform controls → Integrations).
⚡ Run it in one command
No cloud account, no database server, no Docker. SQLite storage, the dashboard, and the 24×7
autonomy loops all run in one process.
git clone https://github.com/strikersam/autonomous-ai-agency.git
cd autonomous-ai-agency
./scripts/quickstart.sh # or: make up
Open http://localhost:8001/ and sign in with the credentials the script prints. The script is
idempotent: it writes a local .env with safe defaults, sets up a virtualenv, builds the
dashboard and launches everything. First run takes a few minutes; after that it starts in seconds.
Want top-tier output for free? Put a free NVIDIA NIM key
(NVIDIA_API_KEY=nvapi-...) in .env and restart. No GPU needed. Docker Compose, Render and
troubleshooting: docs/local-setup.md.
Requirements: Python 3.11+ and Node 20+ (Node only builds the UI; --api-only skips it).
How it works
flowchart LR
U([You / SAM]) -->|goal or URL| CEO[CEO agent]
A[Audit + stack scan] --> CEO
AL[Alerts bell] -->|auto fix tasks| CEO
CEO -->|directives| Q[(Task queue)]
Q --> P[Planner]
P --> E[Executor]
E --> V{Verifier<br/>compile + tests}
V -->|fail| E
V -->|pass| PR[Pull request]
PR -->|you merge| SHIP([Shipped])
V -.lessons.-> P
Q -.outward-facing.-> G{{Your approval}}
G --> P
- Work comes in from you (dashboard, SAM, Telegram, a quick note from your phone), from the audit, from the alerts bell, or from the CEO's own 24×7 assessment cycle.
- The CEO decides, grounded in live state, its own past decisions, and C-suite advice for business questions. It drops directives that keep failing.
- Specialists execute through the plan → execute → verify loop, on whichever of a dozen runtimes fits (internal agent, Hermes, goose, aider, OpenHands, OpenCode, Claude Code, E2B, Docker, and more).
- Every LLM call goes through one router with automatic failover across free and paid providers.
- You merge. Outward-facing work (deploys, auth, secrets) waits for your approval before it even runs.
The full change history is in docs/changelog.md.
🧰 The Agency Harness
Everything that makes the agents behave, packaged so you can use it without running the
agency: 70+ skills, subagents, slash commands, the agent-discipline rules, MCP connector setup
and the autonomous-loop pattern. It lives in harness/ and installs
into any repo:
./harness/install.sh /path/to/your-repo # copies; add --link to symlink
Works with Claude Code, Codex, Cursor and Gemini CLI.
Repository map
| Path | What's there |
|---|
harness/ | Reusable, open-source-ready agent harness: skills, subagents, commands, rules, connectors |
proxy.py, handlers/, router/, packages/ | The AI gateway: OpenAI / Anthropic / Ollama surfaces, auth, model routing, provider failover |
agent/, agents/, services/, tasks/, workflow/, runtimes/ | The agency: CEO, specialists, plan → execute → verify loop, task queue, agent runtimes |
backend/, frontend/ | Dashboard API (FastAPI) and React dashboard |
config/ | Model catalogue, providers, agent policy, sandbox profiles |
loops/, .github/workflows/ | The autonomous loop fleet and its registry |
docs/ | Guides, runbooks, ADRs (index) |
scripts/, tests/, evals/ | Tooling, the pytest suite, agent evals |
deploy/, docker/, Dockerfile*, render.yaml | Deployment (Helm, Docker, Render, Cloudflare) |
scripts/local/ | Local launchers (start_server.sh / .ps1, proxy, Ollama, tunnels, Claude Code setup); install.sh / install.ps1 stay at the root |
What it does
| Area | Highlights | More |
|---|
| AI gateway | Drop-in OpenAI /v1/*, Anthropic /v1/messages and Ollama /api/*; task-aware routing with failover across free and paid providers; token quotas, PII/secret sanitiser and honest per-call cost tracking | model-routing.md · api-surfaces.md |
| The agency | URL → onboarding, a CEO agent with a C-suite advisory board, specialists on a dozen runtimes, self-healing from its own errors, zero-key web reach, WSJF portfolio | platform-guide.md · features.md |
| Audit engine | Deterministic SEO / GEO / AIO / security audit with a 0–100 score, priced findings, a PDF report and Screaming-Frog-compatible CSVs | seo-audit.md · proof/audits/ |
| Interfaces | React dashboard, SAM voice, Telegram, GitHub, and any OpenAI-compatible client including Claude Code | claude-code-setup.md · telegram-bot.md |
Try the audit on its own, no signup and read-only:
pip install -r requirements.txt
PYTHONPATH=. python scripts/run_seo_audit.py --website-url https://yourcompany.com --output-dir ./my-audit
Autonomy with a leash
Nothing merges, deploys or messages externally without your approval. On top of that:
| Guardrail | Where |
|---|
| Kill switch halts all autonomous work, mid-run included | AGENCY_KILL_SWITCH |
| Per-agent daily caps in dollars and tokens | AGENT_DAILY_USD_CAP · AGENT_DAILY_KTOKENS_CAP |
| Canary credential that blocks and alerts on a leak | AGENCY_CANARY_ENABLED |
| Policy engine over tools, filesystem, network, shell, GitHub, MCP and more | config/agent_policy.yaml |
| Approval gates with a TTL; expiry denies | packages/governance/approvals.py |
| Hardened sandboxes: no capabilities, non-root, read-only rootfs, no network by default | config/sandbox_profiles.yaml |
| Audit trail with agent identity on every action, secrets redacted | packages/governance/ |
Governance ships in observe mode: rules are evaluated and logged, and nothing is blocked until
you switch to enforcement. Guide and threat model: docs/governance/.
Models. The free sandbox runs on free-tier models with failover; it shows the loop end to
end, not production quality. Production runs on your keys (Claude, OpenAI-compatible, Bedrock,
NVIDIA NIM, Groq, Cerebras, DeepSeek, local Ollama) through the same router. No data leaves
your server.
Don't trust it, check the proof
Most "autonomous agent" projects show you a demo video. Here are artifacts instead:
| Proof | What it shows |
|---|
| This repo is maintained by its own agents | A large share of the merged pull requests here were opened by AI agent sessions: self-healing systems, provider failover, CI hardening, releases. It's the public commit history, and one GitHub search verifies it. |
| Real audit output | The audit engine run against this project's own site. Findings, scores and the agent delegation plan are committed unedited, including our own imperfect score. |
| 24-hour live sandbox | Onboard any site, watch specialists get provisioned, talk to the CEO agent. It resets every 24 hours. No signup wall. |
Screens
Regenerate with python scripts/capture_screens.py, then python scripts/sync_readme_gallery.py.
💬 Chat — unified assistant
Talk to the CEO agent directly; it decomposes goals and routes work to the right specialists.
📊 Dashboard — system overview
Live agent health, recent activity, and system metrics at a glance.
🗂 Tasks — job lifecycle board
Every AI job made visible: waiting, running, blocked, in review, or done.
🤖 Agents — autonomous team
Your specialist roster — each with its own model, runtime, specialty, and guardrails.
🗓 Schedules — autopilot jobs
Recurring and scheduled autonomous work.
⚡ Skills — agentic capabilities
Reusable runtime skills bound to specialists.
🎯 Portfolio — WSJF roadmap
Prioritised initiatives, sprints, and agile health.
📈 Intelligence — trends & competitors
Trend and competitor signals scoped to each onboarded company's stack.
📚 Knowledge — docs & sources
Wiki pages, source material, and reusable context — your team's memory.
🔌 Providers — models, Ollama & MCP
Connect free/cloud/local AI sources and choose which models are available.
🔭 Logs — traces & observability
Every LLM call: tokens, latency, cost, and decision context.
🐙 GitHub — repos & PRs
Connect repositories and manage the agent's delivery surface.
🏢 Company — operating context
An onboarded company's detected stack, systems, and SEO/health.
✨ Onboarding — setup wizard
Scan a website and stand up its specialist agency in minutes.
🔁 Loops — autonomous fleet
Every autonomous loop catalogued: readiness score, maturity, self-heal coverage, drift status, and cost estimate.
🩺 Doctor — diagnostics
System self-checks and autonomy readiness probes.
🛡 Admin — users & access
Manage users, roles, instance activation, and onboarding gates.
📱 Mobile
Responsive layout — sign in, view the dashboard, and work the task board from a phone.
Architecture and security
A React SPA (Cloudflare Worker) over a FastAPI backend (Render or Docker), MongoDB or SQLite
storage, one LLM router for every model call, and a persisted workflow state machine with
human-in-the-loop gates. Secrets are env-only, every endpoint is authenticated with
three-role RBAC, tasks run in per-task git worktrees, and CI runs Bandit, CodeQL, secret
scanning, dependency CVE audits and SBOM generation. Details:
platform-guide.md ·
threat model.
MIT. See LICENSE.
Need assistance?
The software is free and always will be (MIT). If you'd like a hand putting it to work, or advice on AI adoption in general, I'm happy to help.
I built this platform end to end: multi-provider LLM routing with failover, multi-agent orchestration, human-approval workflows, observability, and a CI pipeline where the agents ship their own code (proof). I consult on AI strategy and agentic automation. I can deploy this platform on your infrastructure with top-tier models, onboard your company, and tune the specialist fleet to your stack. A hands-on 2-week pilot is the usual starting point. Early design partners get generous terms in exchange for a public case study.
Contributing
Issues and PRs are welcome. CONTRIBUTING.md covers dev setup, coding standards and the PR checklist. SECURITY.md covers vulnerability disclosure. A large share of merged PRs here are agent-authored, so reading a few recent ones is the fastest way to see the quality bar.
If you'd rather point an agent at a gap than fix it yourself, open an issue. The fleet's inbound-issue triage may pick it up.
Autonomous AI Agency: the AI team that works while you sleep, on a server you own.
For engineers and operators who want the leverage of frontier AI without the cloud bill, the privacy compromise or the headcount.
If this is useful to you, a star helps other people with the same problem find it.
