> ## Documentation Index
> Fetch the complete documentation index at: https://docs.orbb.com/llms.txt
> Use this file to discover all available pages before exploring further.

# MCP servers

> Two Model Context Protocol servers: Orbb Research, over the public professional record, and your workspace's own relationship graph.

Orbb runs two MCP servers. They answer different questions from different data, so most people want one of them, not both.

> [card] [Orbb Research](https://docs.orbb.com/mcp/research-tools.md)
>
> `https://mcp.orbb.com/mcp/research`
>
> Who knows whom across the **public professional record**. Look around a person or a company in one call, or plan a research run that finds and ranks every route to them. Signs in with OAuth or an Orbb API key.

> [card] [Your workspace](https://docs.orbb.com/mcp/workspace-tools.md)
>
> `https://mcp.orbb.com/mcp`
>
> Your organization's **own** graph in Orbb — the people in your CRM, mailbox and calendar, who in your team knows them, and the paths to target accounts. OAuth only, for organizations with MCP access turned on.

| | Orbb Research | Your workspace |
| - | - | - |
| Data | The public record: careers, schools, investor roles, LinkedIn comments | Your organization's connected CRM, email, calendar and calls |
| Sign-in | OAuth in a browser, **or** `Authorization: Bearer orbb_sk_…` | OAuth in a browser only |
| Cost | Most tools are free; a research run spends Research Tokens | Free |
| Tools | 13, named `orbbai_*` | 8, named `orbb_*` |

Both speak the Streamable HTTP transport, over `POST`, with no session to keep. Neither URL takes a trailing slash.

## What your agent is told

When it connects, Orbb Research sends your agent a set of instructions. Their substance:

* **Orbb reads the public record and nothing of yours.** Where your CRM, mail or calls would sharpen an answer, the answer says so in an `outside` block, naming which source holds it. None of it is required. Bring what you find back in `ask` on the next call — it changes the order of what comes back, not who is in it.
* **Look first, then plan.** `orbbai_who_knows_them` and `orbbai_who_knows_someone_there` answer in one call, for free. A research run reads everyone around a person and ranks them against your goal; it is planned, priced and agreed before anything starts. See [Research runs](https://docs.orbb.com/mcp/runs.md).
* **Ask what is already held.** A held answer is immediate and free; re-researching it spends tokens to learn nothing new.
* **A refusal is an answer.** A plan that cannot run comes back with `refused` and the one thing missing — usually the goal. Ask your user and call again. Do not invent a goal to get past it: an invented goal produces a ranking that looks measured and is not.
* **Research is slow.** Poll rather than concluding early.

> [!NOTE]
>
> The tools say what they could not check. An empty answer from a part of the record Orbb has not examined is **not** evidence that two people do not know each other. See [Unknown is not no](https://docs.orbb.com/concepts/unknowns.md).

## API or MCP?

Use the **API** when you are writing code that needs a definite shape back: a pipeline, an enrichment job, a scoring model. Use **MCP** when a person is asking questions through an assistant, or an agent is deciding what to look at next — the tools carry the guidance an agent needs to use them well, and the research runs that rank routes against a goal are only reachable here.

> [card] [Connect a client](https://docs.orbb.com/mcp/connect.md)
>
> Claude, ChatGPT, Cursor, Claude Code, or anything that speaks MCP.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.