RoboCo

AI Agents Company - A virtual organization of 25 AI agents + 1 human CEO, designed to operate as a complete software development workforce.

_{▶ Quick Guide to RoboCo: Setup a 25-Agent Autonomous AI Workforce}

_{Watch the full 2:33 walkthrough (.mp4) →}

Warning

RoboCo is early-stage, work-in-progress software (v0). It's under active development, runs in a homelab, and will have rough edges, breaking changes, and bugs. It is not production-ready and the API/database schema are not stable yet. Treat it as a working prototype to explore and build on — please don't expose it to the public internet as-is. Issues and PRs very welcome.

Tip

📚 Full documentation: docs.roboco.tech — install & first run, the company model, a page-by-page panel reference, model providers, the optional subsystems, deployment, and the API.

Overview

RoboCo implements a structured organizational hierarchy with formal communication protocols, task management, and quality controls. The system enables a single human (CEO) to orchestrate complex multi-project development at scale.

CEO (You, the human)
    │
    ├── Intake (on-demand interviewer: chats only with you to draft a task)
    ├── Secretary (on-demand chief-of-staff: reads company state, runs gated directives)
    ├── PR Reviewer (read-only main reviewer: inbound external/fork + internal PRs, and the root→master in-path gate)
    │
    └── Board (3 agents)
         ├── Product Owner
         ├── Head of Marketing
         └── Auditor (silent observer, reports to you)
              │
              └── Main PM (coordinates all cells)
                   │
                   ├── Backend Cell (6 agents: 2 Devs, 1 QA, 1 PM, 1 Documenter, 1 PR Reviewer)
                   ├── Frontend Cell (6 agents: 2 Devs, 1 QA, 1 PM, 1 Documenter, 1 PR Reviewer)
                   └── UX/UI Cell (6 agents: 2 Devs, 1 QA, 1 PM, 1 Documenter, 1 PR Reviewer)

The 25 agents = Intake + Secretary + PR Reviewer + the Board (3) + Main PM + the three 6-agent cells (18). Agents run on Anthropic Claude by default, or on xAI Grok, OpenAI Codex, Google Gemini, or Moonshot Kimi K3 (each on its own official CLI and subscription, no metered API key) — see the provider note under Configuration.

How it works

You hand a task to the company; it runs through a real build → review → document → merge pipeline and comes back to you to approve.

One full loop, put simply:

You give the Board a task — they review it. The Product Owner and Head of Marketing turn your ask into requirements and acceptance criteria.
You approve — the Main PM starts the work. A notification asks for your Approve & Start decision; approve, and the Main PM breaks it into per-cell subtasks.
Each cell's PM delegates, supports, and triages its developers (UX/UI, Frontend, Backend).
Developers build it, QA verifies and gates it, Documenters keep the books.
Cell PMs merge their PRs into the Main PM's branch.
The Main PM opens the final PR and notifies you "It's done!" — you approve and merge, or send it back for rework. (Only you ever merge to master.)

— Full circle —

See the full walkthrough, with screenshots →

Or watch the full panel walkthrough (video) →

Project Structure

roboco/
├── roboco/                      # Main Python package
│   ├── api/                     # FastAPI routes & schemas
│   │   ├── routes/              # API endpoints (tasks, git, agents, etc.)
│   │   └── schemas/             # Pydantic request/response models
│   ├── services/                # Business logic services
│   │   ├── task.py              # Task lifecycle management
│   │   ├── workspace.py         # Multi-agent workspace management
│   │   ├── messaging.py         # Agent communication
│   │   └── optimal.py           # RAG/Knowledge base (in-house pgvector)
│   ├── models/                  # Pydantic domain models
│   ├── db/                      # SQLAlchemy ORM & session
│   ├── enforcement/             # Task lifecycle state machine
│   ├── runtime/                 # Orchestrator for agent spawning
│   ├── agents/                  # Agent base classes
│   ├── mcp/                     # MCP server implementations
│   └── config.py                # Application configuration
├── agents/
│   └── prompts/                 # Agent system prompts (roles, teams, identities)
├── docs/
│   ├── rag/                     # Agent knowledge base (indexed into RAG)
│   └── map/                     # Exhaustive codebase map (agent-facing)
├── alembic/                     # Database migrations
├── motion/                      # Video/motion-graphics toolchain
├── panel/                       # Next.js 16 control panel (served on :3000 via nginx)
├── scripts/                     # Bootstrap and utility scripts (bootstrap.sh for make quickstart)
├── CLAUDE.md                    # Claude Code guidance
├── docker-compose.yml           # Full stack, built from source
└── docker-compose.registry.yml  # Full stack, pulled from the image registry

Running RoboCo

You need Docker + Docker Compose and a Claude Code auth directory on the host (~/.claude, mounted into the orchestrator so agents can reach the model). Copy .env.example to .env and set at least ROBOCO_ENCRYPTION_KEY and ROBOCO_AGENT_AUTH_SECRET (that file shows how to generate each). However you start it, the whole company is reachable at one origin: http://localhost:3000.

Optional — run agents on xAI Grok instead of Claude. RoboCo can spawn agents on xAI's official grok CLI authenticated by a SuperGrok subscription (no metered API key). Run grok login once on the host and point ROBOCO_HOST_GROK_DIR at the resulting ~/.grok so it mounts into Grok agents; the orchestrator keeps the ~6h token refreshed for you. See the Grok block in .env.example (ROBOCO_HOST_GROK_DIR, ROBOCO_GROK_AGENT_IMAGE, ROBOCO_GROK_CLI_MODEL, ROBOCO_GROK_REASONING_EFFORT).

Optional — run agents on Moonshot Kimi K3 instead of Claude. RoboCo can spawn agents on Moonshot's official kimi (kimi-code) CLI authenticated by a Kimi subscription (OAuth device-code login, no metered API key). Run kimi login once on the host; ROBOCO_HOST_KIMI_DIR points at the resulting ~/.kimi-code (optional — defaults sensibly) and mounts read-write, since every Kimi agent shares the host's one rotating credential chain. See the Kimi block in .env.example (ROBOCO_HOST_KIMI_DIR, ROBOCO_KIMI_CLI_MODEL — default kimi-code/k3, kimi-code/kimi-for-coding is the cheaper lever).

Optional: run agents on Nebius Token Factory (NVIDIA Nemotron). RoboCo can route the whole fleet through Nebius Token Factory, an OpenAI-compatible inference API serving 60+ open models (NVIDIA Nemotron, DeepSeek, Qwen, Llama) behind one metered API key. No host login or env var is needed: save the key in the panel under Settings -> AI Routing -> Nebius API key (stored Fernet-encrypted server-side), then click the Nebius mode button. The fleet-wide default model is NVIDIA's open-source Nemotron 3 Super (nvidia/nemotron-3-super-120b-a12b); the searchable picker can narrow it to the rest of the Token Factory catalog. Agents spawn on the opencode CLI with rate-limit parking and per-agent token/cost metering. This is the path judged in the Nebius x NVIDIA hackathon; see HACKATHON.md for what was built for it.

Option 1 — Run the pre-built images (quickest)

Every release publishes all RoboCo images to both the GitHub Container Registry and Docker Hub, so you can run the full stack without building anything. One command brings it up:

git clone https://github.com/rennf93/roboco.git && cd roboco
make quickstart                    # no make? run ./scripts/bootstrap.sh directly

make quickstart (scripts/bootstrap.sh) is idempotent — safe to re-run any time. What it does:

cp .env.example .env              # only if missing — an existing .env is never touched
# ...generates ROBOCO_ENCRYPTION_KEY / ROBOCO_AGENT_AUTH_SECRET / ROBOCO_PANEL_AGENT_TOKEN in place

docker compose -f docker-compose.registry.yml pull
docker compose -f docker-compose.registry.yml up -d

# ...then polls until the stack is genuinely ready (health, migrations, Ollama
# models) and prints a doctor-style summary — or fails loud with what to check.

Note: ROBOCO_PANEL_AGENT_TOKEN is a standing CEO credential — blank it if you later arm cloud auth (see .env.example).

A fresh .env also gets ROBOCO_HOST_PROJECT_DIR and ROBOCO_HOST_DATA_DIR pinned to the checkout location — these are the host-side paths the orchestrator bind-mounts into every spawned agent. If you reuse an existing .env that doesn't set them, quickstart checks that the absence is safe: it only passes silently when the checkout is already at /opt/roboco (the compose default). Anywhere else, it fails loud with the exact line to add (e.g. ROBOCO_HOST_PROJECT_DIR=/your/checkout/path), because an unset var makes Docker create an empty directory at /opt/roboco and every agent spawn dies with IsADirectoryError before reading its system prompt. An explicitly-set value is always trusted as-is — split host/daemon setups (remote or rootless Docker, a bind-mounted checkout) legitimately name paths the script can't see.

Choose the registry and version with two env vars (defaults shown) — set them in .env before running make quickstart:

ROBOCO_REGISTRY=ghcr.io/rennf93   # or docker.io/renzof93
ROBOCO_VERSION=latest             # or a pinned release, e.g. 0.15.0

The orchestrator spawns the matching pre-built agent images on demand — no build toolchain or source compile on your host.

Option 2 — Build from source

The same full stack, built locally from the Dockerfiles instead of pulled:

git clone https://github.com/rennf93/roboco.git && cd roboco
cp .env.example .env              # then edit in your secrets
docker compose up -d              # builds images on first run, then starts everything

Option 3 — Local development (no full stack)

For hacking on the code itself, run only the backing services in Docker and the API on your host. RoboCo's own code requires Python 3.13+ (uv will fetch it if needed):

uv sync
docker compose up -d postgres redis ollama   # backing services only
uv run alembic upgrade head                   # migrate the database
uv run python -m roboco.cli                   # API + orchestrator

# Or just the API without the orchestrator:
uv run uvicorn roboco.api.app:app --reload --host 0.0.0.0 --port 8000

Configuration

Key environment variables (see roboco/config.py for all options). The panel's Settings → Feature Flags card is the authoritative source for the full set of 36 feature flags — toggle them there rather than editing env by hand. The bash block below shows the env-var equivalents of a few representative flags plus settings that have no panel toggle:

# API Server
ROBOCO_HOST=0.0.0.0
ROBOCO_PORT=8000

# Database
ROBOCO_DATABASE_HOST=localhost
ROBOCO_DATABASE_PORT=5432
ROBOCO_DATABASE_NAME=roboco

# Workspaces (Multi-Agent Git)
ROBOCO_WORKSPACES_ROOT=/data/workspaces
ROBOCO_WORKSPACE_AUTO_CLONE=true

# RAG/LLM
ROBOCO_LOCAL_LLM_BASE_URL=http://roboco-ollama:11434/v1
ROBOCO_LOCAL_LLM_MODEL=glm-5.3:cloud

# Feature flags (default-off unless noted; toggle from Settings → Feature Flags)
ROBOCO_CONVENTIONS_ENABLED=false        # per-project architectural conventions standard
ROBOCO_TOOLCHAIN_MATCH_ENABLED=false    # build each target project under its own Python
ROBOCO_OVERLOAD_BREAK_ENABLED=true      # park a provider on a persistent model-API overload
ROBOCO_DOCS_SYNC_ENABLED=false          # docs-divergence sync (release → docs-update task). Default-off; when on, a successful release publish originates one bounded, deduped docs-update task against the roboco-website project.
ROBOCO_DOCS_SYNC_MAX_OPEN_TASKS=3       # rolling cap on concurrently-open docs-sync tasks
ROBOCO_DOCS_SYNC_MAX_PER_CYCLE=1        # max docs-sync tasks originated per publish invocation

# Auditor scheduled sweeps (default 6 hours; 0 disables)
ROBOCO_AUDIT_INTERVAL_SECONDS=21600

The 36 feature flags exposed in the panel fall into eight categories:

Communication: telegram_enabled, telegram_inbound_enabled, x_engine_enabled, x_replies_enabled, x_feature_spotlight_enabled — Telegram notifications and the X (Twitter) account.
Content: video_engine_enabled, video_on_release, video_on_spotlight — the video-generation engine and its triggers.
Governance: provisioning_enabled, strategy_engine_enabled, roadmap_engine_enabled, research_enabled, release_manager_enabled — Board programs and the gated release manager.
Infrastructure: ci_watch_enabled, dep_update_enabled, env_sync_enabled, docs_sync_enabled, gateway_health_enabled, self_heal_enabled, self_heal_originate_enabled, sandbox_db_enabled, rag_auto_update_enabled, transcript_prune_enabled — background loops, sandbox DBs, and workspace health.
Quality: conventions_enabled, org_memory_enabled, fable_mode_enabled, possibilities_matrix_enabled, routing_strict — architectural enforcement, organizational memory, behavioral doctrine, and model-routing strictness.
Budgets: task_budgets_enabled — per-project monthly and per-task cost caps.
Vault: obsidian_vault_enabled, vault_intake_enabled, vault_report_enabled, vault_kb_enabled — Obsidian vault projection, intake, reports, and KB indexing.
PR review: external_pr_enabled, internal_pr_enabled, toolchain_match_enabled — inbound/external PR review, internal PR review, and toolchain matching.

Each flag has a one-line description in the panel card. The remaining twelve Board Programs (Pest Control, Spackle, Scales, Dogfood, Periscope, Megaphone, Mirror, Barfly, War Room, Coroner, Librarian, Sentinel) arm per-program on the dedicated Board Programs page (Business section) rather than the Feature Flags card.

Multi-Agent Workspace Structure

Each agent gets their own git clone for parallel development:

{ROBOCO_WORKSPACES_ROOT}/
└── {project-slug}/
    └── {team}/
        └── {agent-slug}/
            └── [git repository]

Example:
/data/workspaces/roboco/backend/be-dev-1/
/data/workspaces/roboco/backend/be-dev-2/

Task Lifecycle

backlog → pending → claimed → in_progress → verifying → awaiting_qa
    ↓                              ↓              ↓           ↓
cancelled                      blocked      needs_revision   awaiting_documentation
                               paused                              ↓
                                                           awaiting_pm_review
                                                                   ↓
                                                           awaiting_ceo_approval
                                                                   ↓
                                                              completed

Assembled, PR-bearing tasks pass through one extra stage — the in-path PR-review gate — before the PM merges:

in_progress → awaiting_pr_review → awaiting_pm_review
   (submit_up /      (pr_pass)
    submit_root)     (pr_fail → needs_revision)

The cell PM's submit_up (cell→root PR) and the Main PM's submit_root (root→master PR) open the assembled PR and enter the gate; a PR reviewer pr_passes it on to the PM merge or pr_fails it back. Leaf dev tasks (reviewed by QA) and branchless coordination roots skip the gate.

API Endpoints

Domain routes are mounted under /api:

Route Group	Description
`/api/tasks`	Task CRUD, lifecycle, claiming
`/api/agents`	Agent management
`/api/projects`	Project (repo) management
`/api/work-sessions`	Git work session tracking
`/api/git`	Git operations (status, commit, push, PR)
`/api/orchestrator`	Orchestrator / dispatcher status
`/api/kanban`	Kanban board views per team
`/api/dashboard`	Dashboard, metrics, and analytics
`/api/notifications`	Formal notifications (ack-required)
`/api/journals`	Agent journals/reflections
`/api/a2a`	Agent-to-agent direct messaging
`/api/optimal`	RAG/Knowledge base queries
`/api/stream`	Stream processing and permissions
`/api/settings`	Feature flags and settings store
`/api/company-goals`	Company goals, scorecard, brand voice
`/api/research`	Web research (external search/fetch)
`/api/cockpit`	CEO read-only business summary
`/api/release`	Release manager proposals and approval
`/api/secretary`	Secretary chief-of-staff reads and directives
`/api/prompter`	Intake live chat (SSE relay)
`/api/pitches`	Board proposals and CEO approve/provision
`/api/playbooks`	Playbook library curation (approve/reject/archive)
`/api/x`	X (Twitter) post queue and credentials
`/api/video`	Video engine request, pipeline, and approval
`/api/tiktok`	TikTok OAuth credentials (write-only)
`/api/telegram`	Telegram notifications bridge
`/api/roadmap`	Board roadmap cycle approval
`/api/board-programs`	Board Programs registry status and run-now
`/api/pest-control`	Pest Control bug-hunt cycle
`/api/periscope`	Periscope market-research briefs
`/api/coroner`	Coroner postmortem reports
`/api/sentinel`	Sentinel quality-drift reports
`/api/spackle`	Spackle gap-fill cycle
`/api/scales`	Scales portfolio-rebalance cycle
`/api/mirror`	Mirror messaging-fix cycle
`/api/dogfood`	Dogfood friction-fix cycle
`/api/github-app`	GitHub App credentials and repo listing
`/api/products`	Product CRUD
`/api/providers`	AI provider model routing
`/api/docs`	Project documentation management
`/api/usage`	Token usage analytics
`/api/system`	System monitoring (rate limits, etc.)
`/api/auth`	Cloud auth login/logout (mounted only when `ROBOCO_CLOUD_AUTH_ENABLED`)

The agent gateway verbs are served separately under /api/v1/flow/{role}/{verb} (intent verbs) and /api/v1/do (content tools) — see the Agent Gateway.

Development

# Install dev dependencies
uv sync --all-extras

# Run tests
uv run pytest

# Format and lint
uv run ruff format .
uv run ruff check .
uv run mypy roboco/

# Type checking
uv run mypy roboco/

Core Principles

Everything is a task - All work is tracked and documented
No work without a task - Create task record first
No task without acceptance criteria - How do we know it's done?
No closure without documentation - Future agents need context
Communication is constant - Stream reasoning, log everything
The Auditor sees all - Quality monitored silently
CEO approves major changes - Human-in-the-loop for critical decisions

Technology Stack

Layer	Technology
API Framework	FastAPI
Database	PostgreSQL + SQLAlchemy (async)
Vector Store	PostgreSQL + pgvector (in-house engine)
Cache/Queue	Redis
RAG Engine	in-house (asyncpg + pgvector, hybrid retrieval)
Embeddings	qwen3-embedding:0.6b (Ollama)
Local LLM	Ollama (glm-5.3:cloud)
Cloud LLM	Claude API (Anthropic) + xAI Grok (official `grok` CLI, SuperGrok subscription) + OpenAI (official `codex` CLI, ChatGPT subscription) + Google Gemini (official `gemini` CLI, OAuth login) + Moonshot Kimi K3 (official `kimi` CLI, Kimi subscription)
Package Manager	uv

Status

Core Infrastructure (Complete)

In Progress

Full agent autonomy testing

Security

Important

Do not expose RoboCo to the public internet as-is. It is designed to run on a trusted private network (homelab / LAN).

Agent authentication. Requests identify the caller with X-Agent-Id / X-Agent-Role headers. The orchestrator issues each spawned agent an HMAC token (X-Agent-Token, signed with ROBOCO_AGENT_AUTH_SECRET) that binds its id, role and team. Token enforcement is gated by ROBOCO_AGENT_AUTH_REQUIRED:

ROBOCO_AGENT_AUTH_REQUIRED unset/false (default): header-trust mode — the role headers are accepted without a token, so any client that can reach the API may claim any role (including ceo). The API logs a warning at startup in this mode. Acceptable only on a trusted network.
ROBOCO_AGENT_AUTH_REQUIRED=true: every request must carry a valid token; an agent cannot spoof another agent's role. The control panel keeps working because nginx — the only trusted hop between the browser and the API — injects the CEO token (X-Agent-Token) on /api and /ws, so the browser never holds the signing secret. Generate that token with make panel-token and set it as ROBOCO_PANEL_AGENT_TOKEN in .env before enabling secure mode.

WebSocket streams. Token enforcement is currently REST-only. The /ws/* endpoints authenticate by agent_id query param at most and do not yet validate X-Agent-Token, even in secure mode — nginx injects the token so the panel works, but a direct WebSocket connection that bypasses nginx is not rejected. In particular the operator stream /ws/system (rate-limit lifecycle + token-usage snapshots for the dashboard) is unauthenticated. These streams are read-only — no control surface, secrets, or task content — but treat the orchestrator port as trusted-network-only until WebSocket auth lands.

Secrets (the Fernet ROBOCO_ENCRYPTION_KEY, GitHub PATs) live encrypted in the database and in gitignored env files — never in the repo. Per-project git tokens are Fernet-encrypted at rest and never returned by the API.

Nebius x NVIDIA hackathon

RoboCo is entering the Nebius x NVIDIA hackathon (submission window August 26 to October 30, 2026): the whole agent fleet runs on Nebius Token Factory with NVIDIA's open-source Nemotron 3 Super as the default model. HACKATHON.md is the submission notes doc: it isolates exactly what was significantly updated during the window (the Nebius provider integration, the v0.30.0 platform release, the post-release panel fixes) for the Stage One review, and walks through the judge path from make quickstart to a Nemotron-routed fleet.

License

RoboCo is licensed under the GNU Affero General Public License v3.0 (AGPL-3.0). See LICENSE for the full text.

The AGPL's network-use clause (section 13) means that if you run a modified version of RoboCo as a network service, you must make your modified source available to its users. This keeps the project open while preventing closed, hosted re-distributions.

Contributing

Contributions are welcome. All contributors must sign the Contributor License Agreement (CLA.md) — this is automated on your first pull request. See CONTRIBUTING.md for the workflow and why the CLA exists.

Name		Name	Last commit message	Last commit date
.claude/rules		.claude/rules
.github		.github
.roboco		.roboco
agents/prompts		agents/prompts
alembic		alembic
deploy		deploy
docker		docker
docs-redirects		docs-redirects
docs		docs
front		front
motion		motion
panel		panel
roboco		roboco
scripts		scripts
tests		tests
video-renderer		video-renderer
.dockerignore		.dockerignore
.env.example		.env.example
.gitattributes		.gitattributes
.gitignore		.gitignore
.pre-commit-config.yaml		.pre-commit-config.yaml
AGENTS.md		AGENTS.md
CHANGELOG.md		CHANGELOG.md
CLA.md		CLA.md
CLAUDE.md		CLAUDE.md
CODE_OF_CONDUCT.md		CODE_OF_CONDUCT.md
CONTRIBUTING.md		CONTRIBUTING.md
HACKATHON.md		HACKATHON.md
LICENSE		LICENSE
Makefile		Makefile
README.md		README.md
SECURITY.md		SECURITY.md
alembic.ini		alembic.ini
deployment.md		deployment.md
docker-compose.registry.yml		docker-compose.registry.yml
docker-compose.yaml		docker-compose.yaml
docker-compose.yml		docker-compose.yml
pyproject.toml		pyproject.toml
usage.md		usage.md
uv.lock		uv.lock
vulture_whitelist.py		vulture_whitelist.py

Navigation Menu

Uh oh!

Repository files navigation

RoboCo

Overview

How it works

Project Structure

Running RoboCo

Option 1 — Run the pre-built images (quickest)

Option 2 — Build from source

Option 3 — Local development (no full stack)

Configuration

Multi-Agent Workspace Structure

Task Lifecycle

API Endpoints

Development

Core Principles

Technology Stack

Status

Security

Nebius x NVIDIA hackathon

License

Contributing

About

Releases

Sponsor this project

Packages

Used by

Contributors

Languages

Uh oh!

Latest commit

History

Folders and files

Repository files navigation

RoboCo

Overview

How it works

Project Structure

Running RoboCo

Option 1 — Run the pre-built images (quickest)

Option 2 — Build from source

Option 3 — Local development (no full stack)

Configuration

Multi-Agent Workspace Structure

Task Lifecycle

API Endpoints

Development

Core Principles

Technology Stack

Status

Security

Nebius x NVIDIA hackathon

License

Contributing

About

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

Watchers

Forks

Releases

Sponsor this project

Packages

Used by

Contributors

Languages