A coding agent that ships

Say what you want.
Get it live.

Forge plans, writes, builds and deploys a web app from one prompt. When something breaks, it reads the error and fixes it, the way you would.

Plans, builds and deploys in one go.

agent run · step 9 of 9
  1. planhabit tracker · 4 files · Next.js + SQLite
  2. write_fileapp/page.tsx
  3. write_fileapp/api/habits/route.ts
  4. install_packagedate-fns
  5. run_buildType error: 'streak' is possibly undefined
  6. read_filelib/streak.ts · finds the cause
  7. write_filelib/streak.ts · default streak to 0
  8. run_buildcompiled in 6.1s
  9. deployhabits-7f3a.forge.app
Live. One build error, fixed by the agent.9 steps

How it works.

One loop, five moves. The agent keeps going until the app is live, or tells you why it can't be.

  1. 01

    Plans.

    Your idea, as a file plan.

    Forge turns a sentence into pages, routes and data, checked against a schema before a line is written.

  2. 02

    Writes.

    Code, one tool call at a time.

    It reads, writes and installs what it needs, and you can watch every step as it happens.

  3. 03

    Builds.

    In a sandbox, not on trust.

    Every build runs in an isolated container with CPU, memory and time limits.

  4. 04

    Fixes.

    Errors are just more context.

    A failed build goes back to the agent, which reads the error, finds the file and tries again.

  5. 05

    Ships.

    A real link, not a mockup.

    When the build passes, Forge deploys it and hands you a URL you can share.

the whole toolbox →planread_filewrite_fileinstall_packagerun_builddeploy

More than a prompt and a prayer.

Generating code is the easy part. Making it build, remember what you asked for and get better over time is the real work.

Tool calling.

Six small tools with clear contracts. The agent picks the next one; the loop stops on success or a step limit.

Structured outputs.

Plans and file edits are typed. Anything that fails validation is sent back with the error, not guessed at.

Memory.

Each project remembers its files and past decisions, so “make the header blue” knows which header.

Retrieval.

Relevant templates and docs are pulled in by meaning, not keywords, before the agent writes code.

Evals.

Every prompt change runs against a fixed set of app specs: did it build, does it match, how many steps did it take.

Streaming.

Tokens and build logs stream live, so you never stare at a spinner wondering what happened.

Your next idea is one sentence away.

Describe it, watch the agent work, and share the link before your coffee gets cold.

Start building