trello-mcp

[圖片]
trello-mcp

Local Trello MCP server in Go. 48 tools, compact responses, guest boards.

[Release] [CI] [License] [MCP: stdio]

16 cards in one list. Response text, characters. Lower is smaller.

[16 cards in one list, response characters: trello-mcp 6,754, Official 29,141, Node 1.8.1 45,335]

77% less response text than the official server for this call (summary mode). Measured 2026-10-17 on one private work board with 144 cards; characters, not tokens. Full mode: 9,112 (69% fewer). For one card's details the official server returned less: 2,662 vs 2,825. All results and method.

Install

curl -fsSL https://raw.githubusercontent.com/avhn/trello-mcp/main/install.sh | sh

Windows: irm https://raw.githubusercontent.com/avhn/trello-mcp/main/install.ps1 | iex. Checksum-verified.

Or give your agent:

Set up trello-mcp by following https://raw.githubusercontent.com/avhn/trello-mcp/main/docs/setup.md; ask me to run its printed absolute-path auth command myself in a private terminal, then reconnect. Never capture authentication output.

Then: run the printed auth command yourself in a private terminal, register, reconnect. Clients and Claude Desktop: docs/setup.md.

Measured comparison

Measured response text. Chars; independent zero-based row scales.

[Five response comparisons; independent row scales]

Tool definitions. Claude tokens from /context, one session. Node not measured. Lower is smaller.

[Claude definition tokens; Node unmeasured]

Claude Code /context in one session; rows over 1k display-rounded to 0.1k. Node could not be loaded into /context. Separately, tools/list: 20,583 vs 33,766 characters (39% fewer vs Node), not tokens.

One board, one day, one account. trello-mcp at commit cb40466 (v1.0.3); Atlassian hosted https://mcp.trello.com/v1 over OAuth, results captured verbatim from a headless Claude Code session; @delorenj/mcp-server-trello 1.8.1 via npx. Board list figures are characters per returned board over different sets (16 / 15 / 29; Node includes 13 closed boards; the extra trello-mcp board is a guest board in another organization). Card details and comments use one card with two comments. The board is private, so others cannot rerun on the same data. Official results reflect its hosted version on that date. Tool definitions are Claude tokens from /context; chars/4 would give about 5,146 for trello-mcp instead of 8,107.

Aggregate data. Method and harness. Percent fewer = round(100 * (1 - ours/theirs)); per-board values are rounded totals/count.

I use Trello with AI assistants a lot, so I built trello-mcp to make it lighter and easier to use.

A thought: with AI agents, managing context may be one of the main optimizations we need to make now. Every tool definition and response a model reads takes space from the actual task, and as context grows, output quality tends to drop.

Features checked 2026-10-17.

Choose official for zero install, web and mobile, Inbox, Planner or keyword search.

Read-oriented profile

Ten read tools (placement; board_use sets the local default board):

TRELLO_TOOLS=workspace_find,board_find,board_use,list_find,card_find,card_get,comment_find,checklist_get,checklist_items,server_health

Synthetic reference

Synthetic fixtures versus raw Trello JSON for the same objects, regenerated by go test ./internal/bench -update.

Synthetic raw pretty JSON is a reproducible reference, not measured output from another MCP server. Tokens use chars/4.

Docs

Setup · Configuration and security · Tools

License

GPL-3.0-or-later, copyright Mert Dede (MIT up to v0.5.0). Independent Go implementation of the delorenj/mcp-server-trello 1.8.1 (MIT) tool surface, not a runtime dependency. Trello is a trademark of Atlassian. trello-mcp is an independent project, not affiliated with or endorsed by Atlassian.