MCP, the model context protocol, lets a coding agent like Claude Code or Codex talk to your tools. With the Tana MCP connected, that agent can read and write your Tana workspace. This walkthrough goes through five levels, from reading a single node to automating work across Gmail, Slack, and more.
What MCP is and how to set it up [0:00]
The local API and MCP docs walk through setup, and you can paste that page straight into your coding agent to have it do the configuration for you. Two quick checks once you're running: open localhost health to confirm the status is okay, and localhost/docs to see the available local API functions and tools. If you get stuck, Tana runs drop-in MCP help sessions, unrecorded, no set agenda, listed on tana.inc/community.
Level one: read [1:50]
The simplest starting point is reading. Run command + K then copy node link, paste it to your agent, and ask what the node says. Because the link carries the node ID, the agent knows exactly where to look. You can also find nodes by natural language, then walk the structure, the parent of a node, its siblings, and read from there.
Level two: write [2:33]
The MCP also writes. Point the agent at a node and tell it what to add, for example add a task underneath, apply a tag, and fill in some fields. If it gets something wrong, like tagging the item instead of adding a child, just tell it what you meant and it edits and cleans up the earlier mistake.
Level three: repeat [3:35]
To make this repeatable, lean on the foundations agents use outside Tana: an agents.md or claude.md file plus skills. These are the external equivalent of system prompts and command nodes inside Tana. Capture how your workspace is set up in that file, and every new thread already knows your Tana system. OpenAI and Anthropic both publish best practices for writing these, and the agent can even draft its own file for you. For a deeper dive, see the self-driving Tana systems lab on driving Tana with an external agent.
Level four: delegate [4:37]
Now have the agent do real work. Put context on a task, for example a shopping task that should only pull links from Amazon, and add more context in an area the task belongs to. Then ask the agent to run the research, taking the task and area context into account, and it comes back with results that respect those constraints.
Level five: integrate and automate [6:24]
The real payoff is that many other tools speak MCP too, Gmail, Slack, Figma, Intercom, and more. With several connected, you can ask for your most recent email and have the agent file it as a read-later task in Tana under the right node, linked back to the source. Put that on a schedule and the agent can, say, scan your inbox each morning and drop any essays in as tasks. New systems labs exploring use cases like this run roughly every other week, listed on tana.inc/community.