weave-agent-skills

Weave

Stop pushing AI hallucinations to your repo. Review AI-generated plans and code before they hit GitHub.

npx skills add weave-ai-dev/agent-skills

---

The Problem

If you use Claude Code, Codex, or any terminal-based AI agent, you already know these two pain points:

1. You can't review AI plans. Your agent generates a 300-line architecture plan in the CLI. You can't open a thread on line 42. You can't drill into a specific decision. Your only option is to copy-paste snippets back into the terminal and argue. It's slow, lossy, and painful.

2. Your commits are dirty. When the AI writes code, you push it to GitHub just to see the diff. That means AI hallucinations, untested assumptions, and garbage code are now polluting your PR history — visible to your entire team. There's no private review step.

Weave fixes both.

---

What It Does

Weave is a human-agent collaboration bridge. It gives you a Web HUD where you can review and refine your AI agent's work before it becomes real code.

The loop:

1. Plan   → Your AI pushes an architecture plan to Weave
2. Review → You open the web UI, do inline edits, start threaded discussions
3. Approve → Hit approve, your AI pulls the finalized plan and writes the code
4. Review → The AI pushes the code diff back to Weave for private, line-by-line review
5. Ship   → Only clean, human-reviewed code hits your repo

Think of it as a private staging layer between your AI agent and your Git history.

---

Install in 30 Seconds

Works with Claude Code, Cursor, Windsurf, Copilot, Gemini CLI, and 30+ agents.

# Auto-detect your agent
npx skills add weave-ai-dev/agent-skills

# Or specify your agent
npx skills add weave-ai-dev/agent-skills -a claude-code

# Global install (available across all projects)
npx skills add weave-ai-dev/agent-skills -g

Prerequisites:

1. A Weave account → Sign up here
2. That's it. OAuth handles authentication automatically — your tool opens a browser to sign in on first connect.

---

What Your Agent Gets

Once installed, your AI agent has access to these MCP tools:

Tool	What it does
push_plan	Submit an architecture plan for your review
pull_plan	Pull back your reviewed decisions and edits
push_code_review	Submit file diffs for private code review
pull_code_feedback	Retrieve your line-by-line comments
list_documents	Browse and filter project documents
list_versions	Access version history snapshots
use_document	Switch document context
update_project	Update project metadata
release_document	Release session locks

---

How It Looks in Practice

In your terminal:

You: Build me a REST API for user authentication with JWT

Agent: I've pushed an architecture plan to Weave for your review.
       → View it at: https://app.useweave.dev/projects/abc123

       Waiting for your approval...

In the Weave Web HUD:

• Read the full plan with syntax highlighting
• Click any line to start a threaded discussion
• Make inline edits directly
• Hit "Approve" when you're satisfied

Back in your terminal:

Agent: Plan approved. Writing code now...
       ...
       Code complete. I've pushed the diff to Weave for your review.
       → Review at: https://app.useweave.dev/projects/abc123/review

---

Why This Matters

• No more dirty commits. Review AI code privately before it touches your repo.
• No more copy-paste arguments. Discuss specific lines in threaded conversations, not terminal back-and-forth.
• No more blind trust. You see exactly what the AI intends to build before a single line of code is written.
• No workflow disruption. One npx command. Your agent handles the rest. You stay in your flow.

---

Supported Agents

Claude Code · Cursor · Windsurf · GitHub Copilot · Gemini CLI · Aider · Continue · and 30+ more via MCP.

---

Community & Feedback

• Discord: Join the community
• In-terminal feedback: Run /weave feedback to send feedback directly from your workflow
• Issues: GitHub Issues

---

License

MIT

---

Plan. Review. Execute. The bridge between AI intent and human control.