Nerra MCP Server Documentation

Last updated: October 2, 2026

Overview

The Nerra MCP server gives AI assistants such as Claude and ChatGPT access to your company's knowledge in Nerra. The assistant sees only what you can see.

Nerra builds a knowledge library from your company's sources: meetings, documents, messages and work items. Through the MCP server an assistant can:

answer questions about your company, with citations to the source;
look up people, roles, decisions, rules and product facts;
browse library folders and documents;
save a note, rule or decision to the library when you ask it to.

The server follows the Model Context Protocol (Streamable HTTP transport) and uses OAuth 2.1 for sign-in. It is for Nerra customers: you need a Nerra account to connect.

Connecting

Server URL: https://control.nerra.ai/api/agent/mcp

Claude (claude.ai, Claude Desktop):

Open Settings → Connectors and choose Add custom connector.
Name it Nerra and paste the server URL.
Click Connect. You are sent to Nerra to sign in and approve access.

ChatGPT:

Open Settings → Apps & Connectors (developer mode for a custom app) and create a new connector.
Paste the server URL and choose OAuth authentication.
Sign in to Nerra and approve access when asked.

The approval screen. After sign-in, Nerra shows an Approve accessscreen: “<assistant> wants to connect to your Nerra workspace.” On it you can:

give the connection a name, which is shown in your Agents list;
narrow access by access level and access domains (optional; empty means everything you can see at the chosen level);
turn Save knowledge on request on or off; it is on by default for a new connection;
click Approve or Deny.

The assistant can only see what you can see. It never gets more access than your own account.

Technical details for developers:

Item	Value
Transport	MCP Streamable HTTP, JSON-RPC 2.0
Protected resource metadata	`https://control.nerra.ai/.well-known/oauth-protected-resource`
Authorization server metadata	`https://control.nerra.ai/.well-known/oauth-authorization-server`
Authorization / token endpoints	`/oauth/authorize`, `/oauth/token`
Grant types	`authorization_code` with PKCE (S256), `refresh_token`
Client registration	Client ID Metadata Documents (`client_id_metadata_document_supported: true`)
Scope	`agent`
Token	Bearer token in the `Authorization` header

A request without a token gets HTTP 401 with a WWW-Authenticate header that points to the resource metadata.

Tools

The server exposes 11 tools: 10 read-only and 1 that writes (save_knowledge). No tool deletes or changes existing knowledge, and no tool reaches outside your Nerra workspace. Every tool returns only items your account is allowed to see.

Tool	Type	What it does	Required arguments
`query_facts`	Read	Curated, structured company knowledge: features, decisions, people, integrations, norms. The main source for questions about the company.	none
`get_company_context`	Read	A short overview of the company, for broad questions or a cold start.	none
`knowledge_overview`	Read	What the knowledge library contains: layers, categories, counts, last update.	none
`answer_library`	Read	A citation-grounded answer about your company, built from its sources. Only for your company's internal knowledge.	`query`
`query_library`	Read	Raw source excerpts (transcripts, documents, messages) used as supporting evidence and quotes.	`query`
`structured_query`	Read	Exact filters, ranges and ordering over structured records (people, teams, processes, norms, facts).	`entity_type`
`get_library_fact`	Read	One fact by its id, with its evidence.	`fact_id`
`get_entity_neighborhood`	Read	Items linked to one item in the company knowledge graph.	`entity_type`, `entity_id`
`list_library_folders`	Read	Folders and sources in the library.	none
`list_library_documents`	Read	Documents in the library, optionally by source and folder, with the full count.	none
`save_knowledge`	Write	Saves a note, rule, decision or person fact to the library, in the user's own words, only when the user asks.	`title`, `text`, `initiated_by`

All read tools are marked readOnlyHint: true. save_knowledge is marked readOnlyHint: false and destructiveHint: false. All tools are marked openWorldHint: false.

Saving knowledge.The assistant saves the user's words verbatim, without summarising or adding conclusions. It picks a theme (team, customers, product, sales, support, processes, terms, tools, finance, notes). It can also mark the note as a rule (human_rule_fact) or a decision (human_decision_fact). Saved items keep the normal access rules and sensitivity of the library.

Usage examples

The examples below use Tavclaro, a demo restaurant-technology company in Nerra. You do not need to name Nerra in the prompt. The assistant calls Nerra when a question is about your company.

1. Company overview

Prompt: “What does our company knowledge base cover? Give me an overview.”
Tools: knowledge_overview, get_company_context.
Result: categories with counts, for example people and roles, Jira work items, documents and product feature statuses.

2. A person and their role

Prompt: “Who is Hana Kim and what is her role?”
Tool: query_facts.
Result: “Hana Kim is the Integration Lead at Tavclaro”, responsible for POS integrations.

3. Product status

Prompt: “Which of our product features are released, and which are still on the roadmap?”
Tool: query_facts.
Result: released features (POS, QR menu, QR payment, basic analytics) and planned ones with target quarters.

4. Work assigned to a person

Prompt: “Which Jira tasks are assigned to Hana Kim?”
Tools: query_facts; the assistant may also use structured_query or query_library.
Result: the list of POS integration tasks, for example “INT-166 Integrate POS: Mar Feliz”.

5. Saving knowledge

Prompt: “Save this to our company knowledge: the weekly integration sync moves to Thursdays.”
Tool: save_knowledge.
Result: a confirmation that the note was saved to the library, in the user's words.

6. A cited answer

Prompt: “How does our customer onboarding process work? Cite sources.”
Tool: answer_library.
Result: the documented onboarding process with citations to its source document.

When Nerra is not used.General questions are answered by the assistant itself, without Nerra. Examples: “Who is the CEO of Microsoft?”, “Explain the difference between a balance sheet and an income statement”, “Write a job description template for a senior software engineer”.

Permissions, data and privacy

What the assistant can see. A connection acts on behalf of the user who approved it. It sees only knowledge that user can see in Nerra, limited further by the access level and access domains chosen on the approval screen. Company access rules and sensitivity labels (public, internal, confidential) apply to every tool call.

What the assistant can change. Only save_knowledge writes, and only when Save knowledge on request is on for the connection. It adds new items; it never edits or deletes existing knowledge. If saving is off, the tool returns an error and nothing is stored.

Data flow. The assistant sends your question to Nerra. Nerra returns the matching knowledge, which the assistant uses in its reply. The assistant provider handles your conversation under its own terms. Nerra stores saved notes and an audit record of tool calls in your workspace.

Disconnecting. Open Agents in Nerra (https://control.nerra.ai/agents) and click Revoke next to the connection. Access ends immediately and the token stops working. Knowledge already saved stays in the library. You can also remove the connector in Claude or ChatGPT settings.

Limits, troubleshooting and support

Limits

save_knowledge: up to 500 saves per connection per day (UTC). Above that the tool returns ingest_quota_exceeded.
A single tool call times out after 100 seconds and returns upstream_timeout. Retrying usually works.
Search results are paged: query_library returns 5 results by default, query_facts 20, list_library_documents up to 200.

Troubleshooting

Symptom	What to do
The assistant asks you to sign in again	The token expired or was revoked. Reconnect the connector and approve access.
“Access denied” or `scope_exceeds_grant`	The item is outside the access level or domains of this connection. Reconnect and widen access, or ask your workspace admin.
`ingest_not_allowed` when saving	Save knowledge on request is off. Reconnect and turn it on.
The assistant answers without Nerra	Ask about your company explicitly, for example “in our company knowledge…”.
A new tool or description does not appear	Disconnect and connect the connector again so the assistant reloads the tool list.

Support: info@nerra.ai