LEARN / CODING AGENT
A coding agent
that proves its changes.
EvoForge works in real repositories with real tools, and treats a change as unfinished until a test shows it does what it claims.
Built from the kernel up.
EvoForge's agent kernel is its own. The tool surface, the context engine, memory, approvals, the sandbox and orchestration are written in Rust as one system, and models plug in at the top. It is not a window over another vendor's command-line agent, so each of those layers can be improved for the way the agent actually works.
Context is compacted against a budget rather than when a session overflows, attachments are tiered by size, and large tool output is spilled to scratch files, so long multi-step work stays coherent.
Edits that respect structure.
EvoForge reads the repository, searches it, and edits code by syntax tree where that is safer than replacing text, so a rename or a refactor lands where it should. It runs the project's own build, tests and linters in a persistent terminal and reads the results.
The mutation gate
For a bug fix, the change first runs in an isolated worktree. Its regression test must fail before the change and pass after it; a test that was already green proves nothing and is rejected. Every fixed bug leaves its guard test behind, so the codebase only ratchets forward.
Specialists, coordinated.
Large work can be split two ways. A lead agent plans and hands parts to specialists, each with its own tools, skills and brief. Or a swarm claims tasks from a shared board with leases and peer notes. Long goals and scheduled work can run unattended and report back by mail, push or desktop notification.
An open ecosystem.
Basic ships with skills for code review from four angles, pull requests, iOS and Android app debugging, app-store release, website recreation, architecture diagrams, deep research and natural writing, and a skill creator for your own. Plugins and MCP servers (stdio and Streamable HTTP, with OAuth) add tools; every tool is namespaced by its server so names never collide.
All of it runs on the model you choose — Claude, OpenAI, Gemini, DeepSeek, Qwen, Kimi, OpenRouter or a local model.
Computer control
Browser, any desktop app, a persistent shell and remote servers — all behind permissions you set.
Read12 / EVOLUTIONSelf-evolution
How real work becomes memory, skills and tools — and why only changes that survive a gate are kept.
Read02 / STARTGetting started
Install, sign in, connect a model, give it a first task, and unlock a module when you need one.
ReadPoint it at a failing test.
Windows and macOS. Bring your own model key or a local model.