// For agents · software for AI
Agent toolbox
Let your agent (Claude Code, Codex, Cursor…) browse, read documents, run code and remember. Pick by what you want it to do — every tool links to its official source and setup.
Looking for AI apps people use? AI Apps →
693 tools12 capabilities632 open sourcechecked 2026-10-22 · official sources only
01What should your agent be able to do?
All 693 →Pick a capability first, then compare inside the category. Each lists three common starting points.
Let it act
Drive browsers, call services, run commands
Drive browsers & computersOpen pages, click, fill forms, screenshot Common: playwright · browser-use · chrome-devtools-mcp 50 → Call external services (MCP)Let agents call GitHub, docs, databases Common: github-mcp-server · context7 · composio 100 → Run commands & scriptsCLIs and runtimes agents call Common: GitHub CLI (gh) · ripgrep · uv 118 →Give it information
Read docs, pages and code; remember context
Read docs & web pagesTurn PDFs, Office files and pages into clean text Common: MarkItDown · firecrawl · docling 65 → Search & vector storesFind relevant chunks for RAG Common: qdrant · chroma · pgvector 57 → Remember contextRemember preferences and projects across sessions Common: mem0 · claude-mem · hindsight 22 → Understand codebasesSemantic search, call graphs, go-to-definition Common: serena · GitNexus · codebase-memory-mcp 19 →Keep it in control
Isolate execution, guard secrets, control cost
Run code in a sandboxMistakes can’t touch your machine Common: E2B · microsandbox · CubeSandbox 49 → Guard secrets & permissionsKeep .env away from models, block risky commands Common: infisical · gitleaks · vault 39 → Switch models, control costOne API for many models, with budgets Common: litellm · OpenRouter · claude-code-router 35 →Make it better
Reuse skills and rules, follow open protocols
Add skills & rulesAGENTS.md and Skills, write once, reuse Common: anthropics/skills · superpowers · addyosmani/agent-skills 59 → Follow protocols & specsMCP, A2A, AGENTS.md themselves Common: modelcontextprotocol · A2A · agents.md 15 →02Connect it to your agent — safely
Where each coding agent keeps MCP config, rules and Skills, and what to check before you install.
03Starter kits
More stacks →Don’t want to compare one by one? These three kits cover most people’s first needs.
Let your agent research the web
The most common first need: look things up, read docs and pages.
- chrome-devtools-mcp — drive the browser
- firecrawl — pages to clean text
- context7 — up-to-date library docs
Let your agent run code safely
Isolate first, then lock down secrets.
- E2B — cloud sandbox
- microsandbox — local sandbox
- gitleaks — scan secrets before commit
Let your agent know your documents
Convert, index, then add long-term memory.
- MarkItDown — documents to text
- qdrant — vector store
- mem0 — long-term memory
04This week
GitHub OSS radar →Only tools whose primary audience is agents; general developer software is excluded. Popularity is always GitHub stars.
05What counts as “for AI”?
One test: who uses it at run time?
THIS PAGE · AGENT TOOLS
The agent uses it; people only install and configure
Interfaces are CLI, MCP, API/SDK; the agent calls it without a human. E.g. Playwright, MCP servers, E2B sandboxes.
→ AI APPS
AI software people operate directly
It has a UI, chat or IDE. E.g. ChatGPT, Cursor, ComfyUI. Tools used both ways (OpenCode, Ollama) are marked “dual-use”. Go to AI Apps →
About this data · inclusion rules · for AI to read
- Three questions decide “for AI”: does an agent call it at run time, through a CLI/MCP/API, without a person in the loop?
- Only official sites and repositories are linked; unauthorised API resellers and cracked software are never listed.
- “Common” picks and starter kits are editorial choices: officially maintained, widely used, and safe defaults.
- Machine-readable: llms.txt · /api/v1/software.json · /api/v1/agent-clients.json