Todoist MCP: Which One Can Actually Do What

Search for Todoist MCP and you get four different answers. Todoist ships a one-click connector for Claude. Todoist also ships a separate MCP server with a different feature set. Two community servers are still on GitHub with a few hundred stars between them. Every one of those pages tells you how to install its thing, and none of them tells you what the other three can do.

That matters more here than it does for most integrations, because these are your tasks. The gap between them is not cosmetic. One of these setups can delete a project. One of them will silently strip the repeat schedule off a recurring task if you phrase a request slightly wrong. One of them ships 47 tools into your context window before you have asked a single question.

So this is the comparison the vendor has no reason to write. What each option can actually do to your Todoist, what it costs you in context, and which one to pick.

What Todoist MCP means in 2026

The Model Context Protocol, or MCP, is the open standard that lets an AI client call outside tools during a conversation. Anthropic introduced it in late 2024, and it is now how most software exposes itself to agents. If you want the fuller version, we wrote a plain English explainer on what MCP is.

Todoist was early to it. The Doist/todoist-mcp repository has been public long enough to collect 536 stars, and unlike a lot of vendor MCP projects it is still being pushed to daily. The description now reads “a set of tools to connect to AI agents,” with MCP as one delivery route rather than the whole product, which tells you where Doist thinks this is going.

The confusion is that the same company also ships a Claude connector, and the two are not the same thing. The connector is a curated, safer subset. The MCP is the full surface. Most guides treat them as one integration with two install methods. They are not.

The four ways to give an agent your Todoist

Here is the state of play, checked against the repositories and the vendor’s own help pages rather than against other people’s tutorials.

Both community servers are genuinely good pieces of work and both still run. The greirson one consolidates related actions into single tools, so todoist_task handles create, get, update, delete, complete and reopen rather than shipping six separate definitions. That is a smarter design than the official server’s, and I will come back to why.

They are still not what I would install today. Neither uses OAuth, so you are pasting a long-lived API token into a config file. Neither has a vendor behind it. And the official server now covers everything they cover and more.

What the Claude connector can and cannot do

This is the part that is genuinely hard to find, so here it is plainly. According to Todoist’s own connector documentation the connector lets Claude read tasks and projects, create tasks and projects, update them, and complete them.

That is the whole list. What it explicitly cannot do:

• Delete anything. No tasks, no projects.
• Manage sections, labels, comments or filters. If your system runs on labels, the connector is blind to half of it.
• Assign tasks in shared projects. Solo use only, in practice.
• Move a project into a team workspace folder, or reorder projects in the sidebar.

There is also one documented behavior worth committing to memory. Ask the connector to “change the date” on a recurring task and it removes the repeat schedule. Ask it to “reschedule” and the recurrence survives. That is a one-word difference between adjusting a routine and deleting one, and nothing in the interface warns you.

My honest read is that the connector’s limits are mostly correct. An agent that cannot delete your tasks is an agent that cannot have a bad day at your expense. If everything you want is “add these seven things to my inbox” and “move that to Friday,” install the connector and stop reading.

Setting up the official Todoist MCP

If you need labels, comments, filters, reminders or delete, you need the server. It is remote, so there is no npm package to install and nothing to keep updated.

For Claude Code, one command.

claude mcp add --transport http todoist https://ai.todoist.net/mcp

Todoist also publishes a Claude Code plugin route, which is the path it recommends.

/plugin marketplace add doist/todoist-mcp
/plugin install todoist@doist

For anything else that speaks streamable HTTP, point it at the same URL. If your client only handles local servers, the package runs standalone.

npx @doist/todoist-mcp

Auth is OAuth and it runs in your browser the first time the agent touches a Todoist tool. Nothing to paste, nothing to export, no token sitting in a plaintext config. That removes the single most common MCP setup failure, which is a key that never made it into the environment the client actually launched from.

Two limits to know before you get ambitious. The server handles a maximum of 25 tasks per call, so a brain dump of forty items becomes two calls. And projects cannot be created directly inside a workspace folder, they have to be moved afterward.

The 47-tool problem nobody mentions

Now the uncomfortable part.

I counted the tool files in the official repository and there are 47 of them. add-tasks, find-tasks, update-tasks, complete-tasks, uncomplete-tasks, reschedule-tasks and delete-object are all separate definitions, and that is before the full set each for projects, sections, labels, comments, filters and reminders, plus get-productivity-stats, get-project-health, analyze-project-health, get-workspace-insights and export-project-template.

It is impressively complete. It is also a lot to hand a model that just wants to add a task.

Every tool definition is text in your context window before the conversation starts. Anthropic’s own engineering write-up on advanced tool use puts real numbers on this. A five-server setup with 58 tools consumes roughly 55,000 tokens before anyone types anything, with GitHub’s 35 tools alone accounting for about 26,000 of it. Anthropic says it has seen tool definitions reach 134,000 tokens before optimization.

Todoist ships 47. I have not metered it myself, so treat this as an order of magnitude rather than a measurement, but on Anthropic’s published ratios a 47-tool server is not a rounding error. It is a meaningful slice of the window you wanted for your actual work.

This is why greirson/mcp-todoist consolidating six task actions into one todoist_task tool is the better design, even though I am telling you not to install it. The official server chose granularity, and granularity has a bill.

The good news is that the fix is arriving on the client side rather than the server side. Anthropic’s Tool Search Tool keeps the tool index outside the context window and loads only the definitions the model actually reaches for, cutting definition tokens by around 85% while improving tool-selection accuracy from 79.5% to 88.1% on Opus 4.5. If your client supports deferred tool loading, turn it on before you connect anything this large.

Until then, the practical advice is boring and works: do not run every MCP server you own at once. Connect Todoist in the sessions where you are planning, and turn it off in the sessions where you are coding.

Do the community servers still make sense?

Briefly, for two cases.

You want the Sync API. stanislavlysenko0912/todoist-mcp-server covers Todoist’s Sync API alongside REST, and it injects your projects and labels into the server instructions so the model does not burn a tool call discovering them. That is a thoughtful touch the official server does not copy.

You want bulk operations and local control. greirson/mcp-todoist does bulk create, update, delete and complete through consolidated tools, with in-memory caching. It runs on your machine against your own token.

Both are worth a look if one of those is a hard requirement. For everyone else the calculus is simple. A first-party server with OAuth, daily commits and a company’s name on it beats a community server that was last touched three months ago, and it is not close. The same logic applies to picking any MCP server: check the last commit date before you check the feature list.

What it costs

The MCP server is free. Todoist’s plan limits are what actually bite.

The free Beginner plan allows 5 personal projects. That sounds fine until an agent with add-projects available decides your brain dump deserves its own project structure. Todoist’s Pro plan raises that to 300 personal projects, and Business adds team projects on top.

Claude Code itself needs a paid Claude plan, which Todoist’s setup article states directly. The connector route has no such requirement.

The thing nobody budgets for is context, not money. A 47-tool server running in every session costs you window on every turn, forever. That is a recurring cost with no line item.

Where ContextBolt fits

Worth being straight about the seam here, because it is a real one rather than a pitch. ContextBolt has nothing to do with Todoist and does not integrate with it.

The two sit either side of the same gap. Todoist holds what you decided to do. ContextBolt holds what you decided was worth keeping, the threads, posts and articles you saved from X, Reddit and LinkedIn because they seemed useful later. It tags and clusters them automatically, and Pro ships an MCP endpoint so Claude can search that pile in the same conversation where it reads your task list.

There is a design point in here too. ContextBolt’s bookmark server ships five tools, not 47. search_bookmarks, list_clusters, get_cluster_bookmarks, get_recent_bookmarks, save_bookmark. That was a deliberate call, made for exactly the reason above. A server you run in every session has to be cheap to have running.

Pro is $6 a month. The free Basic tier gives you 150 bookmarks, which is plenty to find out whether the idea holds.

Worth doing?

Yes, with a caveat about which one.

Start with the connector. It takes one click, it covers add, update, complete and reschedule, and it physically cannot delete your work. Live with it for a fortnight. Most people find that is the whole job, because the thing they actually wanted was to stop opening a second app to capture a task.

Move to the official MCP server when you hit a wall the connector built on purpose. Usually that is labels or filters, sometimes it is cleanup work that needs delete.

And go in knowing what you are handing over. An agent with write access to your task list can reorganize it faster than you can read the diff, which is worth thinking about before you approve the OAuth prompt rather than after. We covered how to vet an MCP server in more detail, and the short version is that first-party plus OAuth plus a recent commit history is about as good as this ecosystem currently gets.

Todoist clears all three. Most of what you will find searching for this does not.