Skip to main content

Docana MCP

Docana's hosted MCP server connects your AI client to your workforce. Create and improve employees, search enterprise knowledge, and manage work in Docana from the conversation you already use. The employees and their configuration stay in Docana, so your team can also use them through connected channels and the API.

Claude, ChatGPT, Cursor, and other clients can connect when they support remote MCP servers. The tools available in a conversation depend on the client, its settings, and your Docana permissions.

Use the connection to:

  • Create and improve employees. Describe a workflow, validate the agent specification, and import it into an application. See Employees.
  • Work with your enterprise context. Search documents, organize collections, and review conversations from your AI client.
  • Keep one home for the work. Continue in Docana, connect channels for your team, or call the same workflows from your own code.
Looking for the other direction?

This page is about using Docana from your AI tools. To give your Docana agents tools from external MCP servers, see MCP Servers.

Connect​

The server lives at https://mcp.docana.com/mcp. There are two ways in:

  • Sign in with your Docana account. For claude.ai and Claude Desktop: add the connector, sign in, done. No key to create or paste.
  • Use a Docana API key. Send it in the Authorization header with the ApiKey scheme (not Bearer), and make it a Full access key. Right for Claude Code, Cursor, and anything driven by a config file or running headless.

Either way, the connection acts as you: it can do exactly what your account or key can do, nothing more.

Works on claude.ai and in the Claude desktop app:

  1. Open Settings → Connectors.

  2. Click Add custom connector and paste the server URL:

    https://mcp.docana.com/mcp

  3. Click Connect and sign in with your Docana account.

To disconnect later, remove the connector from the same Connectors settings.

Things to ask for​

Once connected, talk to it normally. Your AI tool picks the right Docana capabilities:

  • "Search my knowledge for our refund policy and summarize it."
  • "Create a collection called Contracts and upload this PDF into it."
  • "Create an Expenses table, suggest its columns, and connect it to my invoices collection."
  • "Save this expense to my table, then show its values and source reference."
  • "Upload https://example.com/handbook.pdf to the Onboarding collection."
  • "List my applications and show yesterday's conversation threads for the support app."
  • "Build an agent that triages support tickets by urgency and hands the urgent ones to a human."
  • "Validate this agent spec and tell me what's wrong with it."
  • "Rename document 4231 to signed-contract and archive the old draft."
  • "Register this WhatsApp template payload for application 12 and enable it."
  • "Measure how often the support agent hands off to a human, then A/B test the shorter greeting against the current version."

What your AI tool can do​

CapabilityWhat's in it
ToolsKnowledge, documents, collections, tables, file connectors, conversations, agents, agent Insights and experiments, applications, MCP server registrations, and search. Tools carry read-only or destructive annotations; your client decides how to present confirmations.
ResourcesYour content as attachable context: collections (with their documents), document files, conversation transcripts, and the agent-building guide.
Promptsbuild-agent: the same expertise Docana's own agent builder works with, so your AI tool designs agents like Docana does. upload-document: teaches the right upload path for where a file lives. run-experiment: how to configure Conversation Insights and run an A/B experiment end to end, from forking a candidate to promoting the tested version.

Working with tables​

MCP exposes table creation and settings, column suggestions and editing, document sources, collection connections, extraction, live JSON rows, saved records, and CSV export references. These tools call the same authorized operations as the API and CLI. Document Sites and Conversation Insights use the same table IDs as the Tables workspace.

For example, call listTables, then listTableColumns for the chosen table. Use the returned column IDs to save exact values through saveTableRecord, or load a whole CSV keyed by one of its columns with importTableRows. To make a new table out of a CSV, check the proposed columns with previewTableFromCsv and create it with its rows through createTableFromCsv. Read them back with listTableRows. To keep receiving documents from a collection, use connectTableCollection and inspect listTableWatches for connection state.

Use classification tools to preview rules and reclassify saved values. Use delivery tools to prepare a weekly support report, preview its records, and configure email or webhook delivery. PDF and chart generation have separate preview operations, so you can inspect a report before sending it.

Put path and query arguments at the top level and request fields inside body. See Automate tables for the workflow and the API reference for request fields and MCP examples. Reconnect your MCP client after a deployment if it still shows an older tool list.

Uploading files​

The upload method depends on where the file is available. The upload-document prompt helps your client choose:

  • A URL: Docana downloads it (https:// links, up to 25 MB).
  • Text it already has: sent inline.
  • A local file: sent directly as raw bytes through a one-time upload link (up to 100 MB per request on the hosted platform, and the link reports the exact ceiling), so binaries like PDFs upload cleanly. Bigger files go through the chunked protocol the prompt describes.

Building agents​

You can also build through the employee experience. createEmployee saves an employee application with its goal, portrait, category, and persona. askEmployeePreparation uses the same preparation assistant as the UI and returns a threadId for follow-up messages. Employee knowledge, expectations, performance, and secure website login operations are exposed as tools too. See Employees from code for inputs and the complete mapping.

Ask your connected client to describe the workflow, load the build-agent prompt when supported, validate the specification, and import it into the intended application. Inspect the resulting configuration and run representative test cases before making it available to users. Validation checks the specification; evaluations check behavior.

For clients that cannot run a workflow directly, open it in Docana or use the CLI or API. See Employees for an example and a review checklist.

Troubleshooting​

SymptomFix
Custom connector won't finish connectingThe sign-in step opens a Docana login. Complete it in the account you want the connector to act as, then approve access.
401 on every call (API key)The header scheme is ApiKey, not Bearer: Authorization: ApiKey your_api_key_here.
Tools exist but calls return 403The account lacks access. Widget-only and Routine-only keys are locked to their endpoints. Use a Full access key.
A recently added capability doesn't show upMCP clients cache the tool list per connection. Reconnect the server (or restart the session).

Next steps​