From ticket to merged PR. Forge agents plan in isolated git worktrees, implement across your full stack, run CI, and open real pull requests on GitHub. You review and merge.
This isn't a toy demo. We created a ticket to add a delete button to Forge — while recording. The agent planned, coded across Phoenix backend and React frontend, ran CI, opened a PR, and merged it.
Forge is an Elixir/Phoenix umbrella app where every mission is a supervised GenServer process.
Every mission is a GenServer under a DynamicSupervisor with a Registry for unique naming. 20 agents can run in parallel across different repos. If one crashes, the supervisor restarts it. Others keep running.
Every tool call — read_file, patch_file, run_command — is broadcast on a mission_trace:<id> topic. The React frontend picks it up instantly. No polling. No SSE hacks.
Each mission gets its own git worktree. Agents work on the same repo simultaneously without conflicts. Worktrees map 1:1 to GenServer processes.
The runner doesn't care if it's talking to Claude, Gemini, or Ollama. It's a GenServer that sends messages and handles responses. Swapping models is a config change.
Forge's agent loop handles the entire software delivery lifecycle — with humans in control at every critical step.
Submit a ticket in plain language. Paste a user story, a bug report, or just describe what you need. Forge reads your codebase to understand context.
The AI agent generates a step-by-step implementation plan before writing a line of code. You review and approve — or chat with the agent to refine it.
Forge works in an isolated git worktree — writing files, running tests, fixing errors, and iterating until the implementation is clean.
A pull request is automatically created with a full description. CI runs. When all checks pass, you merge — or Forge auto-merges if you configure it.
MissionTemplate
schema + migration in missions.ex
api_controller.ex
+ route in routeruseMissionTemplates
React hook for fetching & savingNewTicketModal.tsx
with apply logicpriv/repo/seeds.exs
Forge is a complete AI engineering platform — not just a code autocomplete. Every feature is designed for real production workflows.
Two-agent architecture: a planner that reasons about your codebase and a reviewer that validates the output. Agents iterate until the work is correct.
When the review agent approves the work, Forge automatically pushes the branch and opens a pull request with a detailed description — no human needed.
Forge polls your GitHub Actions checks in real-time. The dashboard shows live CI status, and merging is blocked until all checks pass.
Get notified when human approval is needed, PRs are opened, or missions fail — via Slack, email, or custom webhooks. Configurable per org.
Set strategic goals, create projects beneath them, and let AI break down your objectives into a ready-to-run mission backlog in seconds.
Every LLM call is logged with token counts, cost, latency, and model. Set per-mission budget caps. View daily spend trends and model performance in the Intelligence Hub.
Run missions on Claude, Gemini, or local models via Ollama. Compare cost and performance. Switch models per mission without changing your workflow.
Viewer, Engineer, and Admin roles. Engineers create and approve missions. Admins can force-complete, merge, and manage org settings. GitHub OAuth built-in.
Watch agent reasoning live via WebSocket. See the code diff, CI status, token usage, and cost — all in one view. Chat with the agent during planning.
Works inside your existing workflow
Join engineering teams already using Forge to clear their backlogs and ship production-quality code autonomously.