Open-source, self-hosted, BYOK. Meetings become briefs, briefs become tickets, tickets become code your team reviews and ships. Your exec approves from Slack.
$ git clone https://github.com/workforce0/workforce0 $ cd workforce0 && docker compose up
The cycle
Meeting → PRD → Review → Dev → QA → Ship → Memory. The Chief of Staff posts to Slack/WhatsApp/Teams; you approve, pause, or cancel by replying.
How it works
Four steps. Each is observable, auditable, replayable. No black box.
Upload a recording, paste a transcript, or dial in by phone. Whisper + domain-aware prompts turn audio into structured notes.
Chief-of-staff agent drafts a brief, asks clarifying questions, and breaks approved work into steps with roles and skills.
Specialist agents (BA, architect, dev, QA) pull tickets. Dev code-gen runs on your laptop via Claude Code / Cursor — source code doesn't route through our backend.
Summaries post to Slack / WhatsApp / Teams. Your exec approves with a reply. The web UI is for audit — not daily work.
Who it's for
Workforce0 splits setup from day-to-day use. One of you installs it. The rest of your team never touches the web UI.
You're comfortable with Docker, API keys, and env vars. You spin up the stack, paste BYOK keys, wire Jira / Slack / Google once. After that, the product runs itself.
Execs, PMs, operators. They get brief summaries in their comms channel, approve with a reply, and move on. The web UI is audit-only — they rarely open it.
Features
Built from the ground up as self-hosted, provider-agnostic, audit-friendly. Each feature is optional — the parts you don't turn on don't run.
BA agent drafts briefs through multi-model consensus across Claude, GPT, and Gemini. The planner gracefully falls back across providers. One provider down? The system keeps moving.
Use the API keys you already pay for. No pay-per-token reselling, no hidden inference fees, no lock-in.
Slack, Google Chat, WhatsApp — two-way reply-to-approve. Teams is one-way notifications today. The exec never opens the app; approvals happen in the room they already work in.
Visual setup for Jira, Google Chat, Drive, GitHub, Twilio. Get-your-API-key buttons that open the right vendor page.
Caller dials your Twilio number, talks to a local agent, brief lands in your queue. Three provider modes — Pipecat (local Whisper + Ollama + Kokoro), Gemini Live, or OpenAI Realtime — with deterministic fallback.
Full TS/JS AST plus a regex-based Python extractor (Go, Rust, Java on the way). God-nodes feed the planner so decompositions target the parts of the repo that matter.
Prisma middleware enforces tenant and project scopes on every read. Audit UI shows the decision trail for every approval.
The dev agent uses your Claude Code / Cursor subscription via a local daemon. Source code flows to the CLI you already trust — not through our backend.
A vendored library of playbooks and specialists the chief-of-staff can pick from. Add your own without forking.
Works with your stack
Every integration is optional and disabled gracefully when keys aren't set. No 'sign up to see what's supported.'
Plus local models via Ollama, vLLM, LM Studio, or any OpenAI-compatible endpoint.
Why open-source
AI platforms that capture your meetings, your decisions, and your code pipeline are too important to rent. Workforce0 is MIT-licensed software you clone and run.
Transcripts, briefs, approvals, comms history — all live in your Postgres. We never see any of it because we don't run anything.
BYOK means you pay OpenAI / Anthropic / Google directly at their actual rates. No resold tokens, no monthly SaaS tier.
Vanilla Postgres, Redis, Node. Run it on Docker Compose, Fly, Railway, Render, DigitalOcean, or a Hetzner box. We don't own your runway.
Standing on shoulders
Workforce0 is a synthesis, not an original thought. Credit where it's due.
The god-nodes idea, EXTRACTED vs INFERRED edge confidence tags, and the graph-based planner enrichment come from safishamsi/graphify.
The 'never mutate past messages' rule that keeps Anthropic's prompt cache warm comes from NousResearch/hermes-agent.
Our 'Aperture' design system borrows heavily from the visual discipline of Linear, Raycast, and Vercel.
A fresh install on a laptop takes a couple of minutes. First brief approved in Slack: same afternoon.
