Everything you need to ship software with a fleet of autonomous agents — from your first product to a landed stack of reviewed PRs.
Self-hosting? Jump to Self-host.
A Product
is an outcome (not a board). Each product has one or more repos
— a GitHub repo you keep on GitHub (we drive it: clone, branch, commit,
PR, merge) or a local
path for desktop / offline work. Agents run against
the product's primary repo.
Describe an outcome. A planner decomposes it into missions, and the bounded dispatcher fans them out to Builder agents — in parallel, on the BEAM — each in its own worktree. Missions flow through four stages as their runs progress.
The product flow view — missions grouped by derived stage.
Tick “Stack the missions” in the Ship-it box and a goal becomes a graph of dependent PRs. The planner says which missions actually build on which; every mission with no upstream starts at once, and each downstream one is released as its upstreams push their branches. They land upstream-first through the merge queue, and GitHub auto-retargets the rest on merge.
Agents don't merge to production on their own. Completed runs land in the Review Inbox; you review the diff, then the merge queue verifies each one behind a gate and lands it serially (a PR on GitHub, a branch merge locally). Landing is dependency-ordered for stacks.
Skills are reusable SKILL.md playbooks — the knowledge an agent draws on for a kind of work. Browse the library at /skills, enable the ones your org's agents may use, or import any MIT repo from GitHub in a click (admin-only; playbook text only, bundled scripts flagged, never run).
Every seat runs the 80/20 router: routine work runs on local / open models (free, private, via Ollama), and only the hard ~20% escalates to a frontier model like Claude Opus — metered against your per-org budget. Bring your own key, or your own Ollama. See /models.
Agents act through a governed MCP boundary — every tool call is authorized against your policy before it runs. Browse 100+ MIT-licensed MCP servers at /integrations, and bring your own MIT server too.
Governance: each agent credential has a role; role-based policy controls which tools and skills it may invoke, who can trigger a run or a merge, and per-org spend caps — beyond what branch protection covers.
Memory: agents remember decisions and conventions across runs (org, product, and mission scope) — injected into every prompt and captured after each success, so the fleet improves over time.
Agentic Flow is MIT-licensed and portable. Clone it, run it yourself, keep your data and models in your environment.
git clone https://github.com/vinpatel/agenticflow
mix setup
mix phx.server # → http://localhost:4000