接入你的智能体
從發布者安裝。不是 PlugMCP 端點。
遠端 URL 還是本機指令? 兩種安裝方式的差異
One command configures all your AI CLIs. Install, share, and version MCP servers, slash commands, and prompts across Claude Code, Gemini CLI, and Codex CLI.
weave install web-dev # Puppeteer MCP → Claude Code, Gemini CLI, Codex CLI
weave install git-tools # Git MCP server — add alongside, no conflicts
weave remove web-dev # clean undo — your manual edits stay untouched
Every AI CLI has its own configuration format. Setting up MCP servers, slash commands, and system prompts means hand-editing JSON and Markdown files scattered across different directories with different schemas.
There's no way to share your setup with a teammate, version it, or switch between contexts (work vs. personal vs. open source). Every developer starts from scratch.
Weave fixes this.
One command, all CLIs configured:
weave install web-dev
# → Puppeteer MCP server
# configured in Claude Code, Gemini CLI, and Codex CLI
Stack packs — dependencies resolve automatically:
weave install web-dev
weave install postgres
# → PostgreSQL MCP added alongside Puppeteer, no conflicts
Switch contexts instantly:
weave profile create work && weave profile add github -p work
weave profile create oss && weave profile add web-dev -p oss
weave use work # → all CLIs reconfigured for GitHub tools
weave use oss # → switch to web development setup
Discover MCP servers from the official registry:
weave search --mcp filesystem
# MCP Registry results for 'filesystem':
# Filesystem Server
# Package: @modelcontextprotocol/server-filesystem (npm)
# Repository: https://github.com/modelcontextprotocol/servers
Pin packs to a project — teammates get the same setup:
weave install web-dev --project
# → MCP servers written to both ~/.claude.json (user)
# and .mcp.json in this repo (project scope)
weave list
# web-dev v1.0.0 — Web development MCP stack
# Scope: user + project (/Users/dev/my-app)
Safe and reversible — your manual config stays untouched:
weave install web-dev --dry-run # preview what would change — no files written
weave diagnose # detect config drift across all CLIs
weave sync # fix it — reapply your profile
weave remove web-dev # clean undo, manual edits survive
Add community pack sources:
weave tap add acme-corp/packs # register a third-party pack registry
weave install internal-tools # resolves from the tap if not in the official registry
weave tap list # see all registered taps
Opt in to pack-defined hooks:
weave install ci-tools --allow-hooks
# → pack hooks (e.g., pre-commit checks) applied to Claude Code
# hooks are never applied without explicit --allow-hooks consent
Connect to remote MCP servers:
# In pack.toml — HTTP transport with auth headers
[[servers]]
name = "remote-api"
transport = "http"
url = "https://api.example.com/mcp"
[servers.headers]
Authorization = "${API_KEY}" # secret stays in your env, never in config
Create and share your own packs:
weave init my-pack
# → scaffolds pack.toml, prompts/, commands/, skills/, settings/
# edit, test, publish — anyone can `weave install my-pack`
Think of packs like Homebrew formulas for your AI CLI setup — community-maintained, versioned, one-line install.
A pack is a pack.toml manifest bundled with MCP server definitions, slash commands, system prompt fragments, and settings. Packs install into the active profile — a named set of packs for a specific context (work, oss, personal). Create profiles with weave profile create, switch with weave use, and recover from config drift with weave sync.
weave install web-dev
│
├─▶ fetches pack content from the registry
├─▶ resolves transitive dependencies
└─▶ applies to each installed CLI — non-destructively
Claude Code: ~/.claude.json, ~/.claude/settings.json, ~/.claude/commands/,
~/.claude/CLAUDE.md
+ .mcp.json (with --project)
Gemini CLI: ~/.gemini/settings.json, ~/.gemini/GEMINI.md
Codex CLI: ~/.codex/config.toml, ~/.codex/AGENTS.md, ~/.codex/skills/
Each CLI has its own adapter — a thin layer that knows exactly how to read and write that CLI's config format. Adapters never wipe your existing config. They only add, track, and cleanly remove what they own. A weave remove is surgical.
Homebrew (macOS and Linux):
brew install PackWeave/tap/weave
cargo-binstall (installs a pre-built binary, no compiler needed):
cargo binstall packweave
Shell script (macOS and Linux):
curl -fsSL https://raw.githubusercontent.com/PackWeave/weave/main/install.sh | sh
Build from source:
cargo install --git https://github.com/PackWeave/weave
[!NOTE] Weave targets macOS and Linux. Windows is not officially supported or tested in CI.
Try your first pack:
weave install web-dev # installs Puppeteer MCP across all your AI CLIs
weave list # see what's installed
weave remove web-dev # clean undo
A pack is a directory with a pack.toml manifest at its root:
[pack]
name = "web-dev"
version = "1.0.0"
description = "Web development MCP stack"
authors = ["yourname"]
keywords = ["web", "browser", "git"]
min_tool_version = "0.4.0" # optional: minimum weave version required
[[servers]]
name = "puppeteer"
command = "npx"
args = ["-y", "@modelcontextprotocol/server-puppeteer"]
transport = "stdio"
[[servers]]
name = "filesystem"
command = "npx"
args = ["-y", "@modelcontextprotocol/server-filesystem"]
transport = "stdio"
Packs can also declare:
~/.claude/commands/ or ~/.codex/skills/CLAUDE.md / GEMINI.md / AGENTS.md between tagged delimitersmin_tool_version rejects install on older weave versions${VAR} references, never values[!IMPORTANT] Packs never store secret values. Env vars are written as
${MY_API_KEY}references into CLI config files — the actual values come from your shell environment at runtime.
[!TIP] Test your pack locally before publishing:
weave install ./my-pack— idempotent, re-reads files on each run.
See pack.schema.toml for the full annotated schema and docs/PACKS.md for quality guidelines.
Some CLIs read both a user-level config (~/.claude/) and a project-level config (.mcp.json in your repo). By default, weave install only writes to user scope. Pass --project to also write MCP servers to .mcp.json in the current directory:
weave install web-dev --project
[!TIP]
weave removecleans up both user and project scope automatically, regardless of which directory you run it from.
Run weave diagnose to detect this condition automatically:
Profile: default
Packs: 1 installed
web-dev v1.0.0
Claude Code: drifted (server 'puppeteer' (from pack 'web-dev') is tracked but missing from claude.json)
Gemini CLI: ok
Codex CLI: ok
1 issue(s) found. Run `weave sync` to fix.
See docs/ROADMAP.md for full milestones.
v0.5 — security hardening (in progress):
weave export — reverse-engineer your existing CLI setup into a shareable packv0.6 — ecosystem depth:
weave diff / weave doctor — show changes and verify server healthweave diagnose --fix — auto-repair config driftAI assistants working in this repo should read AGENTS.md.
See docs/CONTRIBUTING.md.
Apache 2.0 — Copyright 2026 Brenno Rangel Ferrari. See LICENSE.
A pack manager for AI CLI tools — weave together MCP servers, prompts, commands, and settings
工具
尚無工具清單。遠端伺服器會在首次健康檢查後取得工具清單。
常見問題
在本頁複製適用於 Cursor、Claude Desktop、Claude Code 或 VS Code 的 stdio 設定。它使用發布者的安裝方式,而不是 PlugMCP 代理伺服器。
這筆記錄未列出必要的環境變數。
Weave 透過 stdio 在本機執行。
上次檢查時間:2026-10-28。 A pack manager for AI CLI tools — weave together MCP servers, prompts, commands, and settings