Tana renamed itself in March 2026. The app you have been using for years became Tana Outliner, and a new collaboration product took the plain Tana name. That would be nobody’s problem except for one thing. Both products now ship an official MCP server, the two servers are not the same, and they do not point at the same data.
Search for tana mcp and you get all three options stacked on one page of results, unlabeled. Two official surfaces from two different apps, plus a community server sitting at the top with 40 stars that was archived two days ago.
Install the wrong one and nothing errors. You get a clean connection to a workspace that is empty, or a server that can write to your inbox and cannot read a single node back. So here is the sort, verified this week.
- Two Tana apps, two official servers. Tana hosts a remote one at
https://home.tana.inc/mcp. Tana Outliner runs a local one athttp://localhost:8262/mcp. - Most existing users want the local one. The app formerly called Tana is Tana Outliner now.
- The local server only runs while the desktop app is open. No app, no MCP, no scheduled agents. A hosted server has no such catch, and ContextBolt reaches the web pages and X, Reddit and LinkedIn saves Tana never sees, from the same agent.
tim-mcdonnell/tana-mcpwas archived on August 20, 2026. It ran on the write-only Input API, so it could never read your notes anyway.- Writes on the remote server arrive as proposals, not edits. You review them like Tana’s own AI.
Why there are two official Tana MCP servers
The Model Context Protocol, or MCP, is the open standard that lets an AI client call outside tools mid-conversation. If you want the ground-up version first, we wrote a plain English explainer on what MCP is.
The split is the part that catches people. In March 2026 Tana announced a second product, a knowledge graph built for teams and agentic work. The new one took the main Tana brand because it targets a broader audience. The outliner everyone already knew became Tana Outliner, and the rename rolled out to the desktop and mobile apps.
Your data did not move. Your settings did not change. But the docs did, and so did the URLs. Anything at tana.inc/docs now 308s over to outliner.tana.inc, which is why half the guides you find look like they are describing a different company.
Two products means two engineering teams, two roadmaps, and two MCP servers built at different times for different jobs. Nobody on the search results page for this term tells you that, so here is each one on its own.
Tana’s remote server, at home.tana.inc
This is the hosted one, and it belongs to the newer Tana product.
The endpoint is https://home.tana.inc/mcp. It speaks HTTP and authenticates with OAuth, so there is no API key to generate and nothing to rotate. Tana lists Claude Code with a one-line setup, Claude Desktop through Anthropic’s connectors, and any MCP client that supports HTTP transport and OAuth.
What an agent gets is the full surface. It can search, read and traverse your workspace, and it can create, update, move and delete documents. It reaches tables, types, skills, calendars and meeting context. The server also preloads your identity and the members of your space, so the agent knows who it is acting as before it does anything.
The design decision worth copying is on the write side. Changes do not apply straight to your workspace. A write creates a proposal you review, exactly the way Tana’s own built-in AI works. So the failure mode of an over-eager agent is a queue of suggestions you decline, not a workspace you have to restore.
That is the right default and more MCP servers should do it. Handing an agent delete permissions on your notes and trusting the prompt is not a safety model, which is the argument we made at length in how to vet an MCP server before you connect it.
One caveat, and it is the whole reason this post exists. This server reaches the Tana workspace. If your notes live in Tana Outliner, connecting here gets you a working OAuth flow, a green status light, and nothing to read.
Tana Outliner’s local server, on port 8262
This is the one most people searching for “Tana MCP” actually want, and it is the one the search results rank worst.
It is not a package you install. It ships inside the Tana Outliner desktop app and runs by default when the app opens, serving http://localhost:8262/mcp over HTTP. Your notes never leave the machine unless a client asks for them.
Setup has an easy path and a manual one. The easy path is that the app writes the config for you.
Enable it from the app
Open the OS menu, then Options, then Local API settings, and turn on the Claude Code option. Tana Outliner adds the entry while it is running and removes it when you close the app. Confirm it landed before you go any further:
claude mcp listThen run /mcp inside Claude Code, pick the Tana server, and approve the modal that appears in Tana Outliner Desktop. A browser confirmation finishes the flow. Personal API tokens are still supported as a fallback, generated from profile settings under API tokens, and passed as an Authorization: Bearer header over SSE. Use OAuth if you can. A personal token in a config file is a long-lived credential with no expiry, and the OAuth path exists specifically so you do not need one.
The tool set is small and well chosen. Reads cover search_nodes, read_node, list_tags and get_tag_schema, plus listing workspaces and fetching a node’s children. Writes cover import_tana_paste, tag, get_or_create_calendar_node, setting field content and options, creating tags, toggling checkboxes, trashing nodes and renaming them.
get_tag_schema is the one to call first and almost nobody does. Tana is a graph of typed nodes, not a folder of documents. An agent that has read your supertag definitions writes queries against real field names. An agent that has not will guess, and it will guess badly on a workspace built around custom types.
Two requirements bite in practice. The desktop app has to be running, and the workspaces you want have to be loaded in it. Claude Code, Cursor, Codex CLI, Gemini CLI, VS Code and Claude Desktop all connect.
Here is the constraint that follows from all of that, and it is worth deciding on before you build anything on top.
The local server is dead whenever your laptop lid is shut. No scheduled agent, no cron job, no cloud runner and no phone can reach localhost:8262. If your plan was a nightly agent that reviews yesterday’s notes, the local server cannot do it and no amount of config fixes that. Every file-backed note app hits the same wall, which is why the Logseq MCP servers and the Apple Notes ones end up with the same shape of caveat.
The community server was archived on August 20
tim-mcdonnell/tana-mcp is still the first GitHub result for this term. It has 40 stars, it is listed on PulseMCP, mcpmarket, lobehub and glama, and on August 20, 2026 it was archived. The README now says the project is no longer maintained and that the author has moved on from Tana.
Directories rank by stars and update slowly, so it will keep showing up for months. Do not install it.
But the archive date is not the interesting part. This is.
That server could never read your Tana workspace, and it was never going to. It was built on the Tana Input API, and Tana’s own documentation is blunt about what that is. The API is write-only. Read access is not yet available.
Look at what it exposed and the shape is obvious. Twelve tools, and every single one creates or edits something, from create_plain_node, create_formatted_node, create_reference_node, create_date_node, create_url_node, create_checkbox_node, create_file_node and create_node_structure through create_supertag, create_field and add_field_value to set_node_name. There is no search tool in that list because there could not be one.
So for most of the time “Tana MCP” has existed as a search term, the thing it pointed at was a one-way pipe into your inbox. Useful for capture. Completely useless for the thing people actually want, which is asking an agent about notes they already wrote.
The Input API limits are worth knowing anyway, because anything doing capture still runs through them. One call per second per token, a maximum of 100 nodes per call, 5,000 characters per request, and it stops syncing on workspaces over 750,000 nodes.
The one live third-party option
If you want something beyond the two official servers, there is exactly one project worth a look, and it does not appear anywhere on the first page of results.
jcfischer/supertag-cli is a TypeScript CLI with an MCP server bundled in, MIT licensed, 48 stars, last pushed in June 2026. That is more stars than the archived server everyone links to.
It works differently from both official servers. It pulls your workspace out through a browser-automated export, keeps it in sync with delta updates, and builds embeddings over the result so you can search by meaning rather than by string. It ships in three sizes, a full mode with 37 tools, a slim mode with 17 and a lite mode with 21, which is a smarter piece of design than it sounds. It falls back to the Input API for creates and uses the Local API for mutations.
The tradeoffs are real. It needs the Bun runtime, it needs Playwright and Chromium for the export, and it keeps a copy of your graph outside Tana. Pick it if semantic search over your notes is the actual goal. Otherwise the local server is less machinery for the same job.
Which one you want
| Server | Which app | Where it runs | Auth | Can it read? |
|---|---|---|---|---|
| Tana remote | Tana | Hosted | OAuth | Yes, writes land as proposals |
| Tana Outliner local | Tana Outliner | Your machine, app must be open | OAuth or token | Yes |
supertag-cliby jcfischer | Tana Outliner | Your machine, exported copy | API token | Yes, with semantic search |
tana-mcpby tim-mcdonnell | Tana Outliner | Your machine | API token | No. Archived Aug 20, 2026 |
The decision comes down to one question, and it is not a technical one. Which Tana holds your notes?
If it is the desktop outliner you have had open every day for two years, that is Tana Outliner, and the local server on port 8262 is your answer. If you are on the newer collaboration product, or you want an agent that can work while your machine is asleep, the hosted server at home.tana.inc is the one.
Once you are connected, this is the first thing worth asking, because it tells you immediately whether the agent is looking at the right workspace.
Copy this prompt
List my supertags, then show me the schema for the three I use most.
After that, search my workspace for notes tagged with any of them from
the last 14 days and summarize what I was actually working on. Tell me
which supertag fields I leave empty most often.If that comes back with tags you recognize, you installed the right server. If it comes back empty or apologetic, you are connected to the other product.
What none of them reach
Every option on this page has the same blind spot, and it is not Tana’s fault.
They all read what you wrote down in Tana. None of them touch the pile of things you saved because they looked useful later, the articles, the X threads, the Reddit comments, the LinkedIn posts sitting in separate apps with no search worth the name. That material never became a note. It got a tap on a bookmark icon and disappeared.
That gap is the whole reason the “second brain” idea keeps disappointing people. The notes layer gets all the tooling and the capture layer gets none, even though the capture layer is where most of what you found actually lives. We went into that split properly in building a second brain from bookmarks.
Where ContextBolt fits
Straight about the seam. ContextBolt does not integrate with Tana, and it is not trying to replace it.
It sits on the other side of the same gap. Tana holds what you wrote. ContextBolt holds what you saved, the web pages you clipped and the posts and threads you kept from X, Reddit and LinkedIn. A Chrome extension captures them as you save them, tags and clusters them on arrival, and Pro exposes an MCP endpoint so Claude can search that pile by meaning in the same session it reads your Tana graph. The two run side by side in the same agent, one for what you typed and one for what you kept.
There is a shared design instinct here worth naming. ContextBolt’s bookmark server keeps its tool list short, search_bookmarks, list_clusters, get_cluster_bookmarks, get_recent_bookmarks, export_bookmarks and save_bookmark, rather than thirty-seven. Tana Outliner’s local server made the same call with roughly a dozen. A server you leave connected in every session has to be cheap to leave connected, and both teams arrived at that independently.
Basic is free and capped at 150 bookmarks, with AI tagging, topic clusters and instant search on your device, which is enough to find out whether you would use it. Pro is $6 a month, removes the cap, and adds cloud sync and the MCP endpoint with its semantic search.
Worth doing?
Yes, and the setup is genuinely quick once you are pointed at the right server. Fifteen minutes, most of it spent updating the desktop app.
Tana rewards this more than a flat note-taking app does. Because everything is typed, an agent that has read your supertag schemas can ask questions a folder of markdown cannot answer. Which projects have an empty status field. What I tagged with a supertag I created and then forgot. Which meeting notes never produced a task. That is a real capability and it did not exist eight months ago.
Go in knowing the two limits. The local server lives and dies with the desktop app, so nothing scheduled can use it. And the entire history of “Tana MCP” as a search term points at a write-only server that was archived this week, so treat every directory listing you find as stale until you check the push date yourself.
Work out which Tana you are on. Install its server, not the other one. Ask for the schema before you ask for the notes. And while the config is open, put ContextBolt beside it for the half Tana never sees. Basic is free up to 150 bookmarks.