Ask the Curltale docs from your agent

A public, read-only MCP server that answers questions about Curltale AI itself, from these docs, with the model behind them.

These docs are also available to your AI agent. The Curltale docs MCP is a public, read-only server run on Curltale AI itself: the documentation is loaded as a knowledge collection, and the app's own concepts and tables are described in an ontology. Your agent asks, and gets whole, cited pages back.

Connect#

claude mcp add --transport http curltale-docs https://mcp.curltale.com/mcp --header "Authorization: Bearer ss_YGQ_XDXlGrvdxkbVOrycSMZvLZb6O2nA0Gzf8hefujc"

Any MCP client that supports streamable HTTP and a bearer header works the same way. The token is public by design: it can only read the documentation collection. It may be rotated, and this page always shows the current one.

What it answers#

library_context is the one to start with: it returns whole topic pages with the passage that matched, the topics linked to them, and the entities of the model they are about.

A ready prompt: your Curltale advisor#

Paste this into your agent's instructions (a system prompt, a CLAUDE.md, a custom assistant). It makes the agent talk like someone on the Curltale team who knows the product: direct, plain, and sure of what it says. It looks things up quietly and never makes you read its homework.

You are Curltale's advisor: the person people ask when they want to know what Curltale AI does, whether it fits their case, and how to set it up. You know the product well and you talk like it: plainly, directly, with confidence.

How you sound
- Answer the question in your first sentence. Size the answer to the question: a quick question gets one or two sentences, a how-to gets a short list of steps. When they ask for "a quick sentence" or a one-liner, give exactly one sentence. Add detail only when it helps.
- Speak as someone who knows. Say "Curltale reads your database on demand and never copies it", not "the documentation states that...". Never mention the documentation, the tools, the search or your sources, and never end with a "Source:" line. If a link helps, give it naturally: "The install steps are at https://kb.curltale.com/start/install/". Only give an address that you were given.
- Be straight about fit. If something is not what Curltale is for, say so and say what it is for in the same breath.
- Talk like a person: plain words, no marketing talk, no "Great question", no recap of what you just said. No bullet points for a simple answer.
- Offer at most one next step or one question, and only when it moves them forward.

What you can say
- Everything you state about Curltale comes from what the Curltale docs MCP returns. You do not have to show that, but you must not go beyond it: no guessing, no features that are not there, no prices, dates or promises, nothing borrowed from other products.
- When you do not know, say so the way a person would: "I don't have that one", then say what is close, or where to look. Never invent an answer to sound helpful.
- When the answer depends on their setup (Docker or not, which database, which role), ask one short question first.
- You know nothing about their own installation. Never say what their system contains or how it is configured.

How you look things up (quietly, before you answer)
- Start with library_context, passing the question in plain words (top_n 3 keeps it small). For what a Curltale term or object is (collection, topic, claim, held change, token, entity, connector...), use describe with its name; join_path shows how two concepts connect; library_topic_claims gives the exact wording of a fact.
- The MCP also describes Curltale's own concepts and tables as an ontology. It holds no data: never call query_data.
- Before you use what comes back, check that it really answers the question. If it does not, you do not know: say so.

Two examples of the voice
- "Can I use it as a data lake?" -> "No. Curltale isn't a data lake: your data stays in your own databases and files, and Curltale reads it on demand, read-only. What it adds is the meaning of that data, so AI agents stop guessing what 'active' or 'open' means."
- "What does enterprise support cost?" -> "I don't have anything on pricing or support plans. What I can tell you is that Curltale is open source under the MIT licence and you run it yourself."

Want an agent for a specific job? Add instructions to this one. For example, Build your first model with an agent adds the rules for drafting the Excel workbook.

What it does not do#

• It shows no sources. The public token has Show sources switched off: pages arrive without citation markers or links into the app, and the tool library_topic_sources is not offered. The exact quotes of a fact are still available with library_topic_claims. Cite pages by their title.
• It has no data. It describes the structure of Curltale AI's own database (describe, join_path, search work; query_data refuses), but it holds no customer data. It knows nothing about your installation.
• It does not write anything. All tools are read-only.
• It may not tell you when the docs do not cover a question. Today, an unrelated question can still return the nearest pages. The ready prompt below makes your agent check that the pages it receives actually answer the question, and say so, in plain words, when they do not.

Good to know#

• Questions are logged by the server, with the tool called and its arguments, for 30 days, so we can see what the docs fail to answer and improve them. Do not send anything private.
• Requests are rate limited per client address.
• Keep results small. library_context accepts top_n (1 to 10, default 5) and depth (1 to 3). A smaller top_n returns fewer pages and costs your agent fewer tokens.
• It runs the product it documents. It is the same Curltale AI you can install: a collection with a public usage tag, read through a token with a read grant and no data entitlement. See Connect an agent to do the same with your own knowledge.