Trello MCP

Connects Claude (Desktop/Cowork and Claude Code) to a Trello board.

The board is chosen by board_id in your config — nothing about a specific board is hardcoded.

~/mcp-servers/trello-mcp/
    trello_mcp.py     the server — no credentials inside, safe to share
    requirements.txt  mcp==1.27.2 (pinned, see below)
    .venv/            dedicated Python 3.12 environment
~/.config/trello-mcp/config.json   credentials (chmod 600, never shared)

Registered in both:

• ~/Library/Application Support/Claude/claude_desktop_config.json (Desktop/Cowork)
• ~/.claude.json (Claude Code)

Both point at the absolute venv interpreter. That is load-bearing — see below.

---

Why it used to disconnect every morning

The old config launched the server with the bare command python3. There are four python3 binaries on this Mac and only one has the mcp package:

Interpreter	Version	mcp?
/usr/bin/python3	3.9.6	no
/usr/local/bin/python3	3.9.6	no
/Library/Frameworks/.../3.9/bin/python3	3.9.6	no
/opt/anaconda3/bin/python3	3.12.7	yes

Which one python3 means depends on the PATH of the process that launched the app. Launched from Finder or at login, the app inherits the minimal launchd PATH (/usr/bin:/bin:/usr/sbin:/sbin) and gets Python 3.9 → ModuleNotFoundError: No module named 'mcp' → "Trello MCP could not be connected." Launched from a terminal, it inherits the shell PATH where Anaconda comes first → works. Hence: broken most mornings, fine when tested by hand.

Reproduce the old failure and the fix:

env -i PATH=/usr/bin:/bin:/usr/sbin:/sbin python3 "$HOME/Claude/Projects/Daily Trello/trello_mcp.py"   # ModuleNotFoundError
env -i PATH=/usr/bin:/bin:/usr/sbin:/sbin ~/mcp-servers/trello-mcp/.venv/bin/python3 -c "import mcp; print('ok')"  # ok

Naming the interpreter by absolute path means no PATH lookup happens at all. The app launches one known binary, which finds one known package directory. Same result at 7 AM as at 7 PM.

---

What a venv actually is (and what it does not protect against)

Python looks for installed packages in a directory tied to the specific interpreter running. Four interpreters = four separate package directories; mcp was only ever installed into one. A venv creates a private one belonging to nothing but this server.

Technically: python3 -m venv .venv creates .venv/bin/python3 and a pyvenv.cfg recording the base install. On startup the interpreter reads that file and sets sys.prefix to the venv, so import resolves to .venv/lib/python3.12/site-packages.

Honest limitation: on macOS, venv symlinks the interpreter rather than copying it, and the standard library still loads from the base install:

.venv/bin/python3 -> /opt/anaconda3/bin/python3
stdlib             = /opt/anaconda3/lib/python3.12

So this venv still depends on Anaconda being installed. It isolates packages, not the interpreter.

Event	Survives?
Anaconda package updates (conda update)	yes
Anaconda 3.12.x point upgrades	yes
macOS updates	yes
Any PATH change — the actual bug	yes
Shell/profile changes	yes
Uninstalling Anaconda	no — rebuild (below)
Anaconda dropping Python 3.12	no — rebuild
Moving/renaming this folder	no — paths are absolute; update both configs

Rebuild after any of those, ~15 seconds:

cd ~/mcp-servers/trello-mcp && rm -rf .venv
/path/to/any/python3.10+ -m venv .venv && ./.venv/bin/pip install -r requirements.txt

If you ever want true independence from Anaconda, install uv and build the venv from a uv-managed standalone Python. Not done here because it adds a new tool to the machine.

Do not blind-upgrade mcp

requirements.txt pins mcp==1.27.2. This is deliberate: mcp 2.0.0 removed the @app.list_tools() / @app.call_tool() decorator API this server is built on. Installing latest breaks the server instantly with AttributeError: 'Server' object has no attribute 'list_tools' — which looks exactly like the connection failure this whole setup was meant to fix. Upgrade only alongside rewriting the handlers for the new API.

---

How the credentials file works

~/.config/trello-mcp/config.json:

{ "api_key": "...", "token": "...", "board_id": "676f0bfc11723d27f716291d" }

Nothing hosted, no external service — an ordinary JSON file on this disk. Two things make it "secret":

1. It lives outside the code, so copying or publishing trello_mcp.py carries no credentials.
2. chmod 600 — only this macOS user account can read it.

The server reads it at startup and holds the values in memory. To share the server, send trello_mcp.py; the recipient makes their own config with their own Trello key. The value isn't encryption — it's separating the thing you share from the thing you don't.

The Trello token has no expiry (dateExpires: null) and scopes Member/Board/Organization read+write. Unlike the Google Drive server's OAuth refresh token, it does not need periodic re-authorisation.

---

Lists are never hardcoded

Every list is resolved by name at call time against the live API, so a list added in Trello (a new class each quarter) works immediately — no code change, no restart. A 60-second cache keeps that cheap; a cache miss forces one refetch before reporting "not found", so a list created seconds ago still resolves.

Measured on this board: a list fetch is ~307 ms. Without the cache every tool call paid it (get_cards costs two round trips), so a 12-card planning session burned ~3.7 s on refetching alone. With it, one fetch covers the burst.

The only list names in the code are the three workflow anchors at the top of trello_mcp.py:

LIST_TODO    = "To Do Today"
LIST_WORKING = "Working on"
LIST_DONE    = "Done Today"

They're named because the daily routine is defined in terms of them (sort_todo_today can't be told "sort the list that means today" without a name). Rename one in Trello → change it in that block and in the Daily Trello CLAUDE.md. rename_list warns you when you rename an anchor.

---

Tools (37)

Reading — get_all_lists, get_cards (includes attachments, labels, due, position), get_card (full detail), search_cards, get_board_summary

Cards — create_card (with attachments/labels/position), update_card, move_card, complete_card, archive_card, unarchive_card, copy_card, set_card_position, set_card_cover, sort_todo_today

Attachments — get_attachments, add_attachment (URL or upload a local file), view_attachment (returns images viewable — this is how Claude reads a screenshot on a card), delete_attachment

Labels — get_labels, add_label_to_card, remove_label_from_card, create_label

Comments — get_comments, add_comment

Lists — create_list, rename_list, archive_list, move_list

Checklists — get_checklists, create_checklist, add_checklist_item, complete_checklist_item, uncheck_checklist_item, rename_checklist_item, delete_checklist_item, delete_checklist

Known board setting: commenting is disabled

add_comment returns a 401 because the board itself has commenting turned off (prefs.comments: "disabled") — not a token or code problem. To enable: Trello → board menu (···) → Settings → Commenting permissions → Board members. The tool detects this and explains it rather than showing a raw 401.

Note on search_cards

Trello's search index lags a few seconds. A card created moments ago may not appear yet; use get_cards on its list instead.

---

Troubleshooting

# 1. Does the interpreter still work?
~/mcp-servers/trello-mcp/.venv/bin/python3 -c "import mcp; print('ok')"

# 2. Does the server start and reach Trello? (Ctrl-C to exit; silence = healthy)
~/mcp-servers/trello-mcp/.venv/bin/python3 ~/mcp-servers/trello-mcp/trello_mcp.py

# 3. Are the configs still pointing at the venv?
grep -A3 '"trello"' ~/.claude.json "$HOME/Library/Application Support/Claude/claude_desktop_config.json"

After changing either config, fully quit and reopen the app — MCP servers are only launched at startup.

Older copies of this script still sit at ~/Claude/Projects/Daily Trello/trello_mcp.py and ~/Downloads/trello_mcp.py. Nothing points at them; they still contain the embedded credentials, so don't share those two.

---

For the resume / portfolio

Kept here so a future resume session has the honest version rather than reconstructing it.

The story worth telling is the failure, not the feature. This server failed every morning for weeks and worked every time it was tested by hand — which is the whole diagnosis: the code was fine and the environment was not. The config named python3; an app launched at login inherits a minimal PATH (/usr/bin:/bin:/usr/sbin:/sbin) where that resolves to a system Python with none of the required packages, while a terminal inherits a PATH where it resolves correctly. Testing from a shell was the one case guaranteed to pass.

What I actually did: directed the debugging. Mine were the decisions — treating "it works when I test it" as evidence about the test rather than reassurance about the code, and requiring a test that reproduces the real launch conditions (stripped PATH) instead of the convenient ones.

What came out of it, now applied to every integration: absolute interpreter paths in configs, a dedicated environment per project, every dependency pinned, and a doctor script that checks each known failure mode and repairs what it can.

Portfolio-length bullet:

Directed development of three automation integrations connecting macOS to cloud services, running unattended on scheduled jobs. Root-caused a recurring daily failure to PATH resolution differing between shell and app launch contexts, and established absolute-path and pinned-dependency conventions that eliminated it across all three.

Full case study: https://zeno-blade-creator.github.io/projects/personal-integrations.html