# Add to your Claude Code skills
git clone https://github.com/smallchao/EvoForgeEvoForge is an open-source ai agents skill for AI coding assistants such as Claude Code, Codex CLI, and ChatGPT, built by smallchao. 一个高性能的具备进化能力的桌面智能体工作台. It has 51 GitHub stars.
EvoForge's catalog security scan is still queued. You can run an instant dependency and prompt-injection check now with the "Scan for vulnerabilities" button above.
Clone the repository with "git clone https://github.com/smallchao/EvoForge" and add it to your Claude Code skills directory (see the Installation section above).
Yes. SkillsLLM lists many other AI Agents skills you can browse and compare side by side. Open the AI Agents category from the badge at the top of this page, or use the Related Skills and comparison links further down to weigh EvoForge against similar tools.
No comments yet. Be the first to share your thoughts!
Unlocks once the catalog security scan passes (runs nightly).
The deep catalog scan for this skill is still queued. Run an instant dependency check now instead.
This repository hosts EvoForge releases, changelogs, and issue tracking. EvoForge is not an open-source project; nothing in this repository grants an open-source license.
Every agent you've tried has the same flaws: teach it something today, and tomorrow it's gone. It can run commands, but files, terminal, browser, and Git live in five different places. And you don't dare leave the screen during a long task, because you don't know what it might delete.
EvoForge's answer is a loop, not a chat box:
flowchart LR
A([Goal]) --> B([Plan])
B --> C([Execute & Verify<br/><small>browser · terminal · files · git</small>])
C --> D([Deliver])
D --> E([Distill<br/><small>skills · memory</small>])
E -. next time: faster, sharper .-> C
After every task, EvoForge reflects on the run: action sequences worth repeating are distilled into skills, your preferences and project context are written into memory, and a built-in curator periodically merges, dedupes, and prunes what accumulates — so it doesn't bloat with use, it sharpens. The next time a similar task shows up, it reaches for the skill it already earned and skips every pit it already fell into.
$ "ship the landing page"
▸ plan 3 steps, 2 tools required
✓ browse collected 14 sources
✓ terminal build passed, 42 tests green
✓ files wrote src/landing/… (+412 −0)
✦ distilled new skill: deploy-preview-flow
● memory noted: you prefer pnpm
Not "a chat box with tools" — everything a worker needs, in one window:
| 🌐 Browser | Embedded browser panel — the agent scrapes, clicks, and live-casts the page back to you |
| 💻 Terminal | Real shell execution; on Windows, encoding and error streams are handled automatically, output stays clean |
| 📁 Files & Git | Read/write the workspace, generate patches, commit branches, open GitHub PRs directly |
| 🎯 Goal mode | Hand over a long-running goal; it plans, executes, and verifies autonomously within a token budget you set |
| 🧠 Memory & skill panels | Everything it accumulates is visible, editable, deletable — you audit its "experience" |
| 🔍 Full-text search | Every past session is searchable; yesterday's conclusions don't need re-deriving |
The big three, one click each. GitHub, Cloudflare, and Supabase connect through their official CLIs with browser authorization — click, log in, done. More importantly, the agent knows what you've connected. With Supabase linked, ask it to "add an index to this table" and it just does it — no more "I don't have database access."
The long tail, through MCP. Any MCP server — stdio or remote HTTP — works by pasting its config. Servers that require OAuth (Notion, Linear, Sentry, Supabase MCP, and most official remote servers) connect with a single Authorize click: discovery, client registration, and PKCE are fully automatic; tokens persist and refresh on their own.
| Style | Character |
|---|---|
| Cold Gray | The Linear-inspired default — cool, precise, quiet |
| Paper | Warm ink-on-paper, low glare, easy on long sessions |
| Mist | Nordic blue-gray, calm and atmospheric |
| Graphite | Pure high-contrast monochrome, true-black dark mode for OLED |
Four complete interface styles × five accent colors × light/dark = 40 combinations, every one calibrated for contrast, switchable live from Settings. And the UI speaks six languages: English (default), 简体中文, 日本語, 한국어, Deutsch, Français.
Everything stays on your machine. API keys, conversations, memories, and skills live locally in ~/.evoforge. Nothing leaves except the calls to the model provider you configured.
Execution has boundaries. Three sandbox levels (read-only / workspace-write / full access), mandatory confirmation for destructive operations, hard refusal of dangerous command patterns, and an audit trail for everything it runs. Hand it a long task and go make coffee.
Grab the latest version from GitHub Releases:
| Platform | Download |
|---|---|
| Windows | .exe installer |
| macOS | .dmg image |
After installing, Help → Check for Updates keeps you current.
Found a bug or want a feature? Open an Issue — attaching relevant output from the Logs panel makes diagnosis much faster.