Image
The IDE for engineering 2.0
Frontier results. A tenth of the cost.
In our real-world benchmark, Metaphor on an open-weight model matched a frontier model task for task, at 90% lower cost. It learns your repo as it goes, so later tasks use 40% fewer tokens.
Download for macOS
Apple Silicon · Release notes
90% lower cost, same pass rate 40% fewer tokens with memory Code, ML, design, video
Each block is one part of Metaphor. Pick one to see how it works.
See how each part works
Agree on the plan. Then let it build.
Describe the change in a sentence. Metaphor reads the repo and writes a plan you can edit. In Plan mode it cannot touch your code, so nothing changes until you press Build.
How Agent worksMeasured, not claimed: Metaphor real-world benchmark, October 2026.
- −90%cost, same pass rateAn open-weight model with Metaphor's memory passed every task a frontier model passed.
- −40%tokens per taskSame model, with memory and structured compaction against without.
- 2×faster to green testsSame model, same workspaces, with memory against without.
- −100%repo searches once it knows the repoLessons recalled instead of rediscovered, against a cold start on the same model.
Beyond code
One agent for the model, the code and the launch.
Train and evaluate a model, ship the feature, then cut the launch video, all in the same session. ML, image, voice, video and evaluation agents are built in.
Voice
Narration with timings
Video
Launch videos, end to end
Evaluate
A second opinion on the output
Code
AnthropicOpenAIGeminiDeepSeekQwenKimiGLM
Image
GPT ImageFluxImagenRecraftIdeogramLuma
Voice
ElevenLabsMistralxAIFish AudioCartesiaDeepgram
Your key
OpenRouterVercelCloudflare
Marketplace
Teach it new kinds of work.
Connect any MCP server, or install plugins, skills and agents by asking for them. Every install asks first.
In short
What you get.
| Where it runs | On your machine, next to your code. Commands run locally, sandboxed on macOS and Linux |
|---|---|
| Git | Stage, commit and browse pull requests without leaving the app |
| Agent review | Drafts a PR review with findings you edit; nothing posts to GitHub until you submit |
| Before it edits | Plan mode writes the plan first; nothing changes until you press Build |
| Self-improvement | Keeps lessons that worked, retires stale ones, turns repeat fixes into procedures, and shows each one it uses |
| Token cost | Benchmark: an open-weight model matched a frontier model's pass rate at 90% lower cost; 40% fewer tokens once it knows the repo |
| Beyond code | Agents for ML research, video, image, voice and evaluation; diagrams and mockups beside the chat |
| Shell commands | Run in a sandbox, and ask first unless you've allowed them |
| Keys | Use the Metaphor gateway, or your own OpenRouter, Vercel or Cloudflare key |
| Extending it | MCP servers, plugins, skills and custom agents |
Questions
What is Metaphor?
An agent-first IDE you install on your computer. Describe a change and it reads your repo, writes a plan you can edit, then makes the edits and runs the tests. It keeps lessons from each session, routes each step to a right-sized model, and handles work beyond code such as ML experiments, design mockups and launch videos.
How does it keep token costs down?
In our October 2026 benchmark (three real-world tasks, each scored by its own test suite), Metaphor on an open-weight model passed every task a frontier model passed at 90% lower cost, and used 40% fewer tokens per task than the same model with memory turned off. It is a small sample, and the full report is in the repo under bench/agent/results. Each role has its own model: searching, summaries and titles go to small, cheap models, while a stronger model plans, builds and reviews. Large tool output is pruned from the context once it has been used, and switching the main model mid-session warns you first, because it rebuilds the prompt cache at full price.
Can it do more than write code?
Yes. It ships with agents for ML research (PyTorch, JAX, fine-tuning and benchmark runs), video production, image generation, voiceover and evaluation, and it draws diagrams and mockups on a canvas beside the chat.
Does my code leave my machine?
The agent, your repo and the shell run on your machine. Only the prompt and the files the agent picks for it go to the model, through the Metaphor gateway or straight to your own provider. There is no cloud copy of your repo.
Does it run commands without asking?
No. Commands wait for your approval unless you've allowed them, and a request that touches anything denied is denied. Shell commands run in an OS sandbox: Seatbelt on macOS and bubblewrap on Linux. In Plan mode nothing changes in your repo until you press Build.
Is Metaphor free?
Every account starts with $20 of model credit, good for 90 days. After that, pay for the models you use through the Metaphor gateway, or bring your own OpenRouter, Vercel AI Gateway or Cloudflare key, or run models on your computer with Ollama.
Which models does it use?
Any model on the Metaphor gateway, including image, voice and video models, or the models behind your own key. Routine steps such as searching and simple edits go to small, fast models; a large model plans and reviews.
Does it get better at my repo?
Yes. After a task it saves a few short lessons, such as how the repo runs its tests or what you corrected, and recalls them the next time. Lessons that keep working rank higher, stale or duplicate ones are retired, and a fix it repeats three times becomes a single procedure. You can mark a lesson wrong or forget it.
Which platforms does it run on?
macOS, Windows and Linux.
Pricing details are on the pricing page.
Remember why you liked building? Start there.
For macOS, Windows and Linux.
Download for macOS
Apple Silicon · Release notes