Kangram MCP
Open app
← Blog

Manage tasks from Claude with MCP

Connect a task-management MCP server to Claude Code, create and assign tasks, and keep one backlog shared with your Telegram team.

Kangram 4 min read

Connecting a task-management MCP server to Claude can turn a request such as “create a task for the login bug” into a durable board item your team can inspect. This guide explains the intended Claude Code workflow; verify the current configuration on the Kangram setup page before using it.

TL;DR

  1. Create or choose a Kangram board.
  2. Generate a narrowly scoped credential using the current Kangram setup flow.
  3. Add the verified Kangram server configuration to Claude Code.
  4. Ask Claude to propose a task action, review it, and then allow the tool call.

You’ll need a Kangram board (create one on the web or via @KanGramBot) and Claude Code installed.

Step 1 — Get an API key

In Kangram, open Settings → API keys and generate a key. Copy it — you’ll paste it into the config.

Step 2 — Add the MCP server

Claude Code reads MCP servers from its config. Add the kangram entry:

mcp_servers:
  kangram:
    command: npx
    args:
      - kangram-mcp@latest
    env:
      KANGRAM_API_KEY: your-api-key

Restart Claude Code so it picks up the server. On launch it will show kangram among the connected tools.

Check before running: package names, configuration syntax, and transports change. Confirm this example against the current Kangram setup and Claude Code MCP documentation. A locally started process can still send credentials and tool arguments to its configured backend.

Step 3 — Create your first task from a session

Start a session and just talk:

> Create a task: fix the login redirect bug in board 15, priority major, assign to ivan
✓ Task #142 created · status: Pending · assignee: @ivan

> What's currently assigned to ivan?
✓ 4 tasks: #138 (In Progress), #140 (Pending), #141 (Pending), #142 (Pending)

> Mark #138 as Done
✓ Task #138 → Done

Claude maps the request to an available tool and structured arguments. Depending on your client settings, you can review the proposed call before it runs.

Step 4 — Keep one backlog with your team

This is where it gets useful. The board Claude writes to is the same board your team sees in Telegram. So:

  • Claude creates #142 during a coding session.
  • A teammate opens @KanGramBot, sees #142 appear, and picks it up.
  • Later, Claude (or the teammate) marks it Done.

No “the AI’s list vs. the team’s list.” One backlog. This is the core idea behind MCP task management, and Claude is just one client on top of it.

Patterns that work well in practice

  • Plan, then create. Ask Claude to propose tasks for a feature, review them, then have it create the approved ones in a batch. Agents over-create if you let them create freely.
  • Reference task numbers in commits. “Fixes #142” — Claude can read the board, so linking code to tasks stays consistent.
  • Use statuses as a handshake. Have Claude move a task to In Progress when it starts and Done when it finishes. The board becomes a live view of what the agent is doing.
  • Filter before you act. list tasks status:In Progress before starting work avoids duplicating an in-flight task.

Troubleshooting

Claude doesn’t see the kangram tools. Confirm the config block is in the file Claude Code actually reads, the API key is set, and you restarted the session. Run /mcp (or your client’s equivalent) to list connected servers.

Tool calls fail with an auth error. The API key is wrong, expired, or lacks access to the board. Regenerate it and confirm the board ID.

Tasks land in the wrong board. Tools take a boardId. If you have several boards, name it explicitly in the request (“in board 15”).

FAQ

Q: Does this work with the Claude desktop app, or only Claude Code? A: MCP support exists across multiple Claude surfaces, but configuration, transports, and approval behavior can differ. Follow the documentation for the specific client you use.

Q: Can Claude create tasks automatically without me asking? A: Only if you instruct it to within a session. It won’t autonomously run in the background — MCP tools are called on-demand by the model.

Q: Is my code/context sent to Kangram? A: No. Kangram receives the tool arguments (task title, board, etc.) and returns the result. Your codebase isn’t part of that exchange.

Next