A workflow engine for AI agent teams. Describe the team in plain English; Forge runs it. Click any feature for the deep dive.
Describe a team in plain English. Forge generates the workflow, runs each persona, validates every handoff with a management agent — across SDLC, research, content, support, ops.
The software-delivery archetype — PM, Architect, Developer, QA, Reviewer — ships reviewed PRs from GitHub issues. One shape among several.
Our designed-and-documented runtime for fully reproducible workflows: replay a run, branch a decision, audit every persona's LLM call. Design complete, implementation not started — today reproducibility comes from snapshot-per-turn.
Learn more →Anthropic, Google, Z.AI, Ollama, MiMo and self-hosted models — plus any OpenAI-compatible endpoint — with priority weights and automatic failover.
Learn more →Every persona's handoff is a commit on your repo. Diff between attempts, full version history, and a signed who-did-what-when record your auditor can read with git log.
Learn more →Every turn becomes an immutable, git-backed snapshot. Roll back to any point and branch into a different path, continue heavy turns in a cloud sandbox with live streaming, and resume on any machine with full context. Offline-first hooks, per-org KMS encryption — and secrets never enter the transcript.
Learn more →Named characters with a role, a character sheet, persistent memory, and XP earned run over run. Not anonymous model calls — a team that improves.
Meet the characters →Schedule on a cron, fire from a webhook, or hand Forge a roadmap and let it pick the next item, run the SDLC, and auto-merge on green.
See the autonomy →Secure-curl substitutes credentials at the egress boundary — plaintext never enters the agent's context, the transcript, or a commit. Scoped, audited, least-privilege.
See the model →Per-run scoped credentials, a throwaway stack to prove each change, and tenant-level worker isolation — with orphans auto-swept.
See isolation →Race two full teams of agents against the same ticket, side by side. Compare their PRs as plain diffs and ship the better one.
See compete mode →Escalations arrive in Telegram with Retry / Resume / Approve / Terminate buttons. Move the run forward without opening the dashboard.
See the controls →PM/Architect/Developer/QA/DevOps is the preset, not the ceiling. Define your own roles, chain them, and run them with the same handoff contracts.
Extend the engine →Per-turn logs and metrics, a unified activity feed across every run, and a Debug toggle that expands the full trace behind any turn.
See inside a run →Agents generate media through one generic API — provider chosen for you, result stored as a clean link. Every hero on this blog was made with it.
See media gen →Install a daemon on hardware you control and it joins your fleet as a runner. Turns execute on your metal, inside your network — with the orchestration still managed for you.
See self-hosted runners →Send a message and a workflow starts. Speak instead of typing and it transcribes first. Reply to something and it knows what you mean. The guess is always confirmed before anything runs.
See chat dispatch →Attach your own MCP servers so agents reach your systems, and register your own skills so they follow your procedures. Per-organisation, additive, no fork required.
See how to extend →An agent joins the meeting, answers when addressed by name, starts a workflow mid-conversation and reports the result out loud — then writes the summary when the call ends.
See it on a call →Create agents, author workflows, change a schedule, ask what's running, undo a mistake — in plain language. Every write is scope-gated, confirmed before it fires, audited and reversible.
See the control plane →A real browser, not an HTTP fetch. Agents open JavaScript-rendered pages, extract readable content, drive multi-step sessions, and hand back artifacts as links.
See how it browses →Roadmap, code, decisions — all on GitHub. Star the repo to follow along.
Follow on GitHub →