Weft
사용하는 모든 AI를 위한 하나의 보드: 에이전트가 작업을 클레임하고, 수락하거나 반려하는 증거와 함께 완료합니다. OAuth를 지원하는 호스팅 MCP 서버 — Claude, ChatGPT, Cursor, Codex, Gemini가 동일한 작업 보드에서 작동합니다. 작업을 클레임하여 두 에이전트가 중복하지 않도록 하고, 진행 상황을 보고하며, 반드시 결정해야 할 사항만 질문하고, 증거와 함께 작업을 반환합니다.
호스팅형 MCP 서버
npx add-mcp 'https://letsweft.com/api/mcp'Claude Code, Codex, Cursor 등에 설치됩니다
문서
Weft MCP Server
One board for every AI you use: agents claim tasks and finish with proof you accept or return.
Weft is a task board for solo founders and small teams who work with AI. One hosted MCP server means VS Code, Claude, ChatGPT, Cursor, Copilot CLI, JetBrains, Kiro, Codex, Gemini CLI, Warp, goose, Raycast, Perplexity, Grok, Le Chat and a dozen more all work from the same board — no extra plugins needed. Thirty-three of them have a copy-paste setup guide.
What makes it a board for agents rather than a to-do list with an API: an agent claims a task with a lease, reports progress to keep it, and finishes with a receipt — artifacts you can check, and what it did not do — which you accept or send back from an Inbox, on the web or your phone. When it needs a decision only you can make, the question waits for you instead of being guessed at.
Server
| Endpoint | https://letsweft.com/api/mcp |
| Transport | Streamable HTTP — MCP 2026-07-28, and 2025 revisions for older clients |
| Auth | OAuth 2.1 — Authorization Code + PKCE, dynamic client registration (RFC 7591): a browser window opens to sign in with your Weft account on first use. For clients without OAuth (scripts, CI, some CLIs), a personal access token from Settings → Access tokens, sent as Authorization: Bearer wft_… — see access tokens. |
| Manifest | /.well-known/mcp.json |
| Registry | Published in the official MCP Registry as com.letsweft/weft (domain-verified namespace) |
A free account takes a minute: letsweft.com/sign-up — the free plan includes 50 active tasks and 200 tool calls a month.
Install
Setting this up from inside an agent (Cline, Cursor, a CLI)? Point it at llms-install.md — the same steps, written for a machine.
VS Code — install the
Weft extension
(also on Open VSX, which is where
Cursor, Windsurf and VSCodium look) and the server registers itself. Without an
extension, add to .vscode/mcp.json, or run "MCP: Open User Configuration" to
have it in every workspace:
{ "servers": { "weft": { "type": "http", "url": "https://letsweft.com/api/mcp" } } }
Switch Chat to Agent mode and ask for your board; VS Code opens a browser for OAuth the first time.
GitHub Copilot CLI — run copilot, type /mcp add, choose HTTP, paste the
endpoint, leave headers empty, set tools to *.
Claude Code
claude mcp add --transport http weft https://letsweft.com/api/mcp
Cursor — Add to Cursor, or add to ~/.cursor/mcp.json:
{ "mcpServers": { "weft": { "url": "https://letsweft.com/api/mcp" } } }
Replit — Add to Replit, or Integrations → "MCP Servers for Replit Agent" → "Add MCP server".
ChatGPT — published plugin: Plugins → search "Weft" → Connect.
Codex CLI
codex mcp add weft --url https://letsweft.com/api/mcp
Gemini CLI — official extension:
gemini extensions install https://github.com/AndreiFinogeev/weft-gemini-extension
JetBrains (via Junie) / Kiro / Warp / goose / Raycast / Perplexity / Grok /
Le Chat / Claude Desktop / Lovable / Windsurf / Cline / Zed / anything else — add
https://letsweft.com/api/mcp as a remote MCP server (custom connector) and
complete OAuth. Per-client walkthroughs: letsweft.com/integrations.
Tools (32)
| Category | Tools |
|---|---|
| Search | search, fetch |
| Board | initialize_board, get_board_state, list_columns |
| Tasks | list_tasks, create_task, bulk_create_tasks, update_task, move_task, trash_task, restore_task, archive_done_tasks |
| Doing the work | get_my_work, start_task, report_progress, complete_task |
| Asking the human | request_input, submit_answer |
| Memory of decisions | record_decision, get_context, get_task_history |
| Projects | list_projects, create_project, update_project |
| Sprints | list_sprints, get_active_sprint, create_sprint, start_sprint, complete_sprint, add_task_to_sprint, remove_task_from_sprint |
The four in Doing the work are the ones that make this a board for agents
rather than a to-do list with an API. start_task CLAIMS a task with a lease
and returns a runId and leaseToken; report_progress renews the claim,
because an agent that dies mid-run cannot report anything and silence is the
only signal the board has. complete_task takes a structured receipt —
artifacts a person can check, claims, what was explicitly NOT done — and when
the task carries machine-checkable verification, the server re-runs those
checks itself before the card moves to Done.
request_input is the other half: a question only the person can answer
survives the session it was asked in, waits in their Inbox (web or phone), and
comes back to whichever client asked.
Destructive tools carry MCP safety annotations; deleting is safe —
trash_task keeps tasks recoverable for 30 days. Full reference:
letsweft.com/docs/mcp-tools.
What you can ask
- "Add a high-priority task: fix the flaky CI test."
- "What am I working on right now?"
- "Move the auth refactor to Done."
- "Start a one-week sprint focused on the landing redesign."
- "Organize my board into projects."
Privacy & security
Your board is private: every session is scoped to your account via OAuth, and tokens are issued per client. Tasks created by AI carry creator attribution, so the board always shows who (or what) added each card.
Support
- Docs: letsweft.com/docs
- Email: support@letsweft.com
- Privacy: letsweft.com/privacy