Forge is a workflow orchestrator, not a coding agent. Your AI tools — Claude Code, Codex, Cursor — already write the code; Forge runs the process around them. It gathers requirements, briefs the AI with your architecture and standards, routes each step to the right model, pauses at the checkpoints you choose, and writes every result back to your tracker. Think of it as the engineering lead that makes sure your process is followed — without writing a line itself.
Those tools are great at writing code in a single session — but the quality rides on whoever's prompting, and nothing carries from one engineer to the next. Forge sits a layer above them: every session gets the same context and standards, each step is pinned to the right model, the gates you set are enforced, and everything is recorded. Same tools — a governed, consistent process around them, so output doesn't depend on your best prompter.
No. Forge runs inside Claude Code, Codex, and Cursor — the tools they already use. No new app, no retraining. An engineer kicks off a workflow (e.g. “implement SHI-784”) and keeps working in their tool while Forge handles the process around them. The skills and conventions your team already wrote become steps Forge can use.
It's built to do the opposite. Forge briefs the AI with your codebase and standards up front, so engineers stop re-explaining context and prompt-wrangling every session — and the first pass holds up instead of bouncing back in review. It only pauses where you've decided a human should sign off; everything else runs automatically. Less busywork, not more process.
No. Your existing instructions stay exactly as they are. CLAUDE.md, .cursorrules, and other tool configs keep governing how the AI writes code in a session. Forge operates at the layer above: what the AI is scoped to work on, what context it gets, which checkpoints it must pass, and what gets logged. Your coding conventions are never touched.
Yes. Every workflow is a list of steps you control — each with a model tier, an approval gate, and any required skills. An admin can edit, reorder, or disable any step, set it once for the whole org, or fork it for a single team. You can even describe a new workflow in plain English from inside your AI tool, and Forge shows exactly what changes before it saves.
Every workflow step is pinned to a model tier — Haiku, Sonnet, or Opus. Routine steps like fetching a ticket or posting status run on the cheaper, faster models; only the hard, code-heavy steps reach the most capable one. You get the right result at the right price without anyone switching models by hand.
It works on Claude Code and any tool that supports model switching. On tools that don't, the workflow runs on whichever model you started with.
It reads the usage your tools already record — input, output, and cache tokens — with no proxy between you and the model. It normalizes every model and tool to a common “weighted token” so spend is comparable on one axis, then prices it at published list rates. It's a transparent estimate, and it says so: it excludes flat-rate subscription plans and any negotiated discounts.
No. A step's tier is a target the session matches in both directions — so the hard, code-heavy steps always land on a capable model and the routine ones always land on a cheap one. If your session is running a frontier model, routine steps like fetching a ticket or posting status are handed down to a cheaper one; if you started on a cheaper model, the genuinely hard steps escalate up to a capable one. Quality is never traded for cost.
The hand-off runs in an isolated sub-agent, so the swapped-in model gets a clean, focused context for that one step and returns the same result. It works on Claude Code and any tool that supports sub-agents or model switching; on tools that don't, your session stays on whichever model you started with.
AI tools: Claude Code, Codex, and Cursor each have a dedicated plugin, plus anything that supports MCP — the open Model Context Protocol. Trackers and docs: JIRA, Linear, and Notion connect through your AI tool's own integrations, so Forge writes work back to whatever you already use. It also runs with no integrations at all — every step still works, and anything it would write to a tracker is handed to you to paste.
Sign up, install the plugin for your AI tool — it takes seconds — and trigger your first workflow. Forge is free for individuals, with every core feature included. Teams of up to 10 add shared workflows, roles and permissions, and org-wide analytics, free during beta; team pricing to be announced at launch.
Yes. Forge orchestrates by sending text instructions to your AI tool — it never touches your repository, CI/CD, or production, and it can't read or write your files. It never receives your API keys, passwords, or tokens (your tool handles those), and it never trains any model on your data. What it sees is structured workflow metadata — which steps ran, timing, token usage — not your raw conversations. Full detail is on our Privacy page.