Skip to main content
Glama

Trellis

An MCP server that reads a Trello workspace and explains how the work is actually flowing.

Trello tells you what is on the board. Trellis tells you what that means: which column work is piling up in, which cards have not moved in two weeks, what changed since Monday, and what is overdue and unowned. Every capability is exposed as a real Model Context Protocol tool, so the same server works in Claude Code, Claude Desktop, or any other MCP-compatible client.

Design note: no second LLM

The analysis tools do not call a language model. They compute counts, ages and outliers from the Trello API and return structured metrics, along with the formula used to rank anything. The client's model reads those numbers and does the reasoning.

That is deliberate. It means the server needs one credential instead of two, it stays cheap and fast, its output is reproducible, and the model can openly disagree with a heuristic instead of inheriting a verdict it cannot inspect.

Related MCP server: Trello MCP Server

Tools

Data

Tool

What it does

list_boards

Every board the account can open, with ids and URLs. Start here.

get_board_info

One board's metadata: description, visibility, members, last activity.

list_board_lists

The open lists (columns) on a board, with card counts and an inferred stage.

list_cards

Open cards on a board, optionally filtered to one list, as triage digests.

get_card

One card in full: description, labels, due date, checklists, recent comments.

search_trello

Search cards and boards, including Trello operators like label:bug or due:week.

Analysis

Tool

What it does

analyze_board_health

Per-list load plus the overdue, unassigned and stalled cards by name.

detect_bottlenecks

Ranks non-done lists by a volume-and-age pressure score, formula included.

generate_standup_report

What moved, what was created and what was discussed in the last N days.

Writes

These change the real board immediately.

Tool

What it does

create_card

Add a card to a named list, with optional description and due date.

move_card

Move a card to another list on the same board.

update_card

Change name, description, due date, due-complete, or archive the card.

add_comment

Post a comment on a card as the authenticated user.

Boards can be given by id or by name, so analyze_board_health("Sprint Board") works without looking an id up first.

Setup

1. Get Trello credentials

  1. Go to trello.com/power-ups/admin and create a Power-Up (any name; it exists only to give you an API key).

  2. Open it and copy the API key.

  3. Next to the key, click the Token link, authorize, and copy the token.

The token is tied to your Trello account and grants exactly the access you approve. Treat it like a password: it is not in git, and .env is gitignored.

2. Install

python -m venv .venv && .venv/Scripts/pip install -e .

On macOS or Linux use .venv/bin/pip instead.

3. Configure

cp .env.example .env

Then fill in TRELLO_API_KEY and TRELLO_TOKEN.

Running it

Local smoke test

.venv/Scripts/python scripts/smoke_test.py

This hits the real Trello API read-only and prints what it finds, which is the fastest way to confirm your credentials work before wiring up a client.

In Claude Code

claude mcp add trellis -e TRELLO_API_KEY=your_key -e TRELLO_TOKEN=your_token -- /absolute/path/to/.venv/Scripts/python.exe -m trellis.server

In Claude Desktop

Add this to claude_desktop_config.json, then restart Claude Desktop:

{
  "mcpServers": {
    "trellis": {
      "command": "C:\\absolute\\path\\to\\.venv\\Scripts\\python.exe",
      "args": ["-m", "trellis.server"],
      "env": {
        "TRELLO_API_KEY": "your_key",
        "TRELLO_TOKEN": "your_token"
      }
    }
  }
}

With the MCP Inspector

npx @modelcontextprotocol/inspector .venv/Scripts/python.exe -m trellis.server

Over HTTP

PORT=8080 .venv/Scripts/python -m trellis.smithery_app

Serves the MCP streamable HTTP endpoint at /mcp. This is the mode a hosted deployment uses.

Deploying to Smithery

The repo ships smithery.yaml and a Dockerfile for Smithery's container runtime.

  1. Push this repo to GitHub.

  2. At smithery.ai, click Deploy and connect the repo.

  3. Smithery builds the image, sets PORT, and proxies to /mcp.

Users then supply trelloApiKey and trelloToken in Smithery's config UI, which is declared by the configSchema in smithery.yaml.

How credentials reach the server

Hosted, one process serves many people, so credentials cannot come from the environment. Smithery passes each caller's config as base64-encoded JSON in a config query parameter on every request to /mcp.

smithery_app.py decodes that and binds the credentials to a ContextVar, not a module-level global. This matters: with a global, two users hitting the server at the same time will overwrite each other's token, and requests get made with the wrong person's credentials. tests/test_smithery_config.py has a test that interleaves two requests and fails if their tokens cross.

Environment variables still work and are used when no request config is present, so the same build runs locally over stdio and hosted over HTTP.

Two notes if you adapt the Smithery Python cookbook

  • Its example stores the API key in a module-level global. Fine for a demo, wrong for anything multi-tenant.

  • It rewrites /mcp to /mcp/, which was correct for FastMCP 1.x. MCP SDK 2.x mounts the route at /mcp, and Starlette's redirect_slashes sends /mcp/ back to /mcp, so that rewrite causes an infinite redirect loop. This repo normalises the other way.

Things worth knowing

  • Rate limits. Trello allows 300 requests per 10 seconds per API key. The analysis tools use three or four calls each, so normal use is nowhere near the ceiling, but a 429 is reported back in plain language if you hit one.

  • Stage detection is name-based. A list is called done, blocked, in-flight or backlog by matching words in its name. A board with unusual column names will have its stages mislabeled, and the card counts are still correct regardless.

  • 404 means invisible, not absent. Trello returns 404 rather than 403 for a board you cannot see, so a 404 may mean the token lacks access rather than a bad id.

  • Archived cards are excluded from board reads. update_card(archive=true) closes a card rather than deleting it; nothing in this server deletes anything.

Example prompts

Once it is connected, these all work in plain language:

  • "Which board has the most stalled work?"

  • "Give me a standup update for the Sprint board covering the last 3 days."

  • "What is overdue and unassigned across my boards?"

  • "Where is work piling up on the Engineering board, and does the data actually support that?"

  • "Add a card to Backlog called 'Rotate Trello token' due next Friday."

License

MIT

Available Tools

13 tools
add_commentA

Post a comment on a card as the authenticated user. This writes to the real board.

Inputs: text (required), card_id (required). Returns: object

analyze_board_healthB

Assess a board: per-list load, overdue cards, unassigned work and stalled cards.

Returns counts and the specific offending cards, not a verdict. Read the numbers and say what they imply, including when the board is simply small.

Inputs: board (required), stale_after_days. Returns: object

create_cardA

Create a card on a board, in the named list. This writes to the real board.

due takes an ISO-8601 timestamp such as 2026-09-01T17:00:00Z.

Inputs: name (required), board (required), list_name (required), due, +1 more. Returns: object

detect_bottlenecksA

Rank the non-done lists by how much work is piling up and ageing inside them.

The pressure score is a crude volume-plus-age heuristic and is returned with its own formula so it can be argued with rather than trusted blindly.

Inputs: board (required). Returns: object

generate_standup_reportA

Summarise what actually happened on a board recently.

Combines card movements, new cards and comments from the activity feed with the board's current overdue and stalled work, which is enough material for a standup update or a weekly status note.

Inputs: board (required), window_days. Returns: object

get_board_infoA

Get metadata for one board: name, description, URL, visibility, last activity.

board accepts a board id or a board name.

Inputs: board (required). Returns: object

get_cardA

Get one card in full: description, labels, due date, checklists and recent comments.

Inputs: card_id (required). Returns: object

list_board_listsA

List the open lists (columns) on a board, in board order, with card counts.

Inputs: board (required). Returns: object

list_boardsA

List every Trello board the authenticated account can open.

Returns board ids, names and URLs. This is the entry point for every other tool, which all need a board id or name.

Inputs: include_closed. Returns: object

list_cardsA

List open cards on a board, optionally filtered to one list by name.

Each card is trimmed to a triage digest: labels, assignee count, due date, overdue flag, days idle, and checklist progress.

Inputs: board (required), limit, list_name. Returns: object

move_cardA

Move a card to a different list on the same board. This writes to the real board.

position is "top", "bottom", or a numeric string.

Inputs: card_id (required), to_list (required), position. Returns: object

search_trelloA

Search cards and boards across the whole account using Trello's own search.

Supports Trello search operators such as label:bug, due:week, is:open and @username.

Inputs: query (required), limit. Returns: object

update_cardA

Update a card's name, description, due date, or archive it. This writes to the real board.

Only the arguments you pass are changed. archive=true closes the card, which hides it from the board but does not delete it.

Inputs: card_id (required), due, name, archive, +2 more. Returns: object

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A3.9/5.0

    Scored across 13 tools

    Disambiguation4/5

    Most tools target distinctly different resources and actions: boards, lists, cards, search, and write operations are clearly separated. The only mild overlap is between analyze_board_health and detect_bottlenecks, both assessing board workflow strain, but their descriptions clarify one is overall health and the other is list pressure ranking.

    Naming Consistency5/5

    Every tool follows a lowercase verb_noun pattern: list_boards, get_card, create_card, move_card, add_comment, and so on. Even the more analytical tools like analyze_board_health and detect_bottlenecks fit the same predictable style.

    Tool Count5/5

    Thirteen tools is within the ideal 3–15 range and every tool earns its place in a Trello-focused server. The set balances read, write, search, and analysis tools without padding or redundant duplication.

    Completeness4/5

    The toolset covers core workflows well: board discovery, list and card inspection, card creation/update/move/comment, archiving via update, search, and higher-level workflow analysis. Missing board/list creation and card deletion are minor gaps for a board-management assistant and can be worked around.

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      D
      maintenance
      Enables interaction with Trello boards, lists, and cards through the Trello REST API. Supports board management, card operations, member management, labels, and checklists through natural language.
      1
      MIT
    • A
      license
      Not graded
      quality
      D
      maintenance
      Provides seamless integration with Trello's API to manage boards, lists, and cards through natural language. It supports full CRUD operations, card movement, and the ability to load Trello resources directly into an LLM's context for analysis.
      MIT