The Kater MCP server gives your own AI agent governed access to your company’s data. Your agent answers questions from saved queries and shared models, builds apps and workspaces, and asks before it changes anything. Kater runs no chat of its own: you keep working in the agent you already use.

Find your address

Open Profile › AI app, or go to /profile#ai. Connect your agent there asks for your AI provider, when your organisation allows more than one, and the app to Open prompts in. Then it shows your organisation’s MCP address, with that app’s own button or command. The address ends in /mcp. Until an agent is connected, Ask covers its box with Connect your agent, and Connect in Profile opens this section.

In a Claude organisation or a ChatGPT workspace, an admin adds Kater for everyone first, and each person then only connects. Admins find the addresses, the AI providers your organisation allows and the steps for each provider on Set up the Kater MCP.

Connect your agent

The status in Profile › AI app follows along: Waiting for Claude, then Claude connected. Ask opens its box as soon as an agent is connected.

Manage connected agents

Open Profile › Agents to see every agent that works with you.

• Share my open pages with my agents. When on, your agent can see your Kater tabs, open a page in one and read the rows on screen.
• Disconnect. Kater then refuses that agent’s calls. Reconnect lets it back in.

Profile › AI app holds your AI provider, the app Kater opens your prompts in, and Connect your agent. See Change your AI app.

Hosted and local servers

The hosted server checks access exactly as the portal does. For several local clients at once, kater serve --transport http serves on 127.0.0.1 behind the token in KATER_MCP_TOKEN.

Tools

The guide vfs://guides/content-layout.md explains the content layout and the safe way to change it. The agent reads it with read_file.

How your agent changes things

• Try first, then ask. With kater, the agent writes only in your private brainstorm folder. Real changes go through kater_apply, which your agent runs only after you say yes.
• Every write can be undone. Each change returns an undo, and History shows it in the portal.
• Some commands stay in your terminal. The agent cannot run login, logout, auth, debug, serve, worker, git, job, init, agent, vscode, scope, parse, retry or execution, and of admin it runs only admin conversations. From new, it makes workspaces, queries, charts, apps, fields, tests and scratch files, and not semantic models, DB datasets or Kater datasets. It can file a promotion request for you.
• Every conversation is audited. Each one is an audit thread. People with audit:read read them on the Conversations page or with kater admin conversations, and query the calls with kater logs audit.

Prompts are your starters

Every starter in your library is also an MCP prompt, so you can start it by name in your agent. In Claude Code, type /mcp__kater__ and the starter’s name.

Each starter is a skill or workflow under a connection’s ai_tools/, or under your own. Read and edit them in AI Tools › Skills and AI Tools › Workflows. See Starters for the full list.

The rules your agent reads first

Your agent calls kater_start first in every conversation. Most clients do not give an MCP server’s instructions to the model, so Kater puts nothing important there. The result of kater_start holds all that the agent needs:

Kater’s rules tell the agent to ground each figure, read the context first, read the guide before it makes an object, use one workspace, try things in brainstorm/ and ask before it changes anything else, search for a measure before it makes one, and compile before it says it is done.

Your own rules

rules.md at the content root holds your organisation’s own rules. When the file has content, context lists it and the agent reads it with read_file. Kater’s rules win on grounding figures and on asking before a change. Your rules win on everything else.

# Our rules

- Call yourself "your data assistant".
- Revenue is net of refunds and tax.
- Use finance's definition when two sources disagree, and say so.

context also lists these files when they have content:

• about.md at the content root, about the company.
• connection_context.md and ai_tools/skills/query-authoring.md of each connection the person can read.
• The person’s own about.md in each of those connections.

Guides and paths

The agent reads a Kater guide with read_file and its vfs://guides/NAME.md path, such as vfs://guides/dashboard.md. The object index names the guide for each object. A path that starts with vfs:// is a file in your content. The guides are in guides/. They are Kater’s, the same for every organisation, and nobody can change them.

Each tool description is at most 2,048 characters, because clients cut it there. The server does not start when a description is longer.

The name your agent shows

The server calls itself Kater, with Kater’s mark, for every organisation. Your agent lists one Kater entry, whichever organisation you sign in to. When two Kater entries are on in one agent, such as two organisations or two stacks, a prompt that Kater writes names the portal it came from, and the agent uses the entry whose kater_start answers that portal. Tool titles name only the action, such as “Run a Command”.

If something looks wrong

• “This Kater has no MCP address set up yet.” Your Kater server has no MCP address. Ask your platform admin.
• The agent’s calls are refused. The agent was disconnected. Select Reconnect in Profile › Agents.
• “Your organisation does not allow Codex with Kater.” Your organisation does not allow that app’s AI provider. Ask an admin to allow it on Admin › MCP. See Set up the Kater MCP.
• The agent cannot see your page. In Profile › Agents, make sure Share my open pages with my agents is on.
• The agent ignores a rule. Check that the rule is in rules.md at the content root, and that the agent called kater_start. Kater’s rules win on grounding figures and on asking before a change.

Next steps

• Connect your agent
• Starters
• Scheduled workflows
• Kater CLI