Skip to main content
Origin’s Model Context Protocol (MCP) server lets you query your endpoint and agent observability data from any MCP-compatible client — Claude Desktop, Claude Code, Cursor, Cline, and others. Ask natural-language questions about your AI agent activity, prompt usage, model spend, and endpoint behavior, and the LLM you’re using translates them into structured queries against your Origin tenant. This page covers what the server does, how to connect, sample prompts, common errors, and how to get support.

​
What the MCP server gives you

Origin’s MCP exposes read-only access to your analytics graph — the same data that powers the Origin console. From an MCP client you can answer questions like:
  • Which agents are running on which endpoints, and what models are they calling?
  • How much are we spending on each model, broken down by department or team?
  • What prompts fail most often, and where?
  • Which sessions had the longest tool-call chains yesterday?
The server is read-only. It cannot modify endpoints, agents, alerts, or any other Origin state.

​
Tools exposed

The server exposes two tools. Most MCP clients discover them automatically once connected.

​
analytics_describe_schema

Returns the full analytics graph: entities, fields, metrics, join edges, the query DSL reference, and worked examples. Takes no parameters. Use this when an LLM needs to learn the shape of your data before composing a query. Most clients call it implicitly the first time you ask a data question.

​
analytics_query

Executes a structured query against your tenant’s data and returns rows plus query cost. Parameters (the LLM fills these in for you): The response includes the result rows, a truncated flag, and a cost object showing how many rows and bytes were scanned, how long the query took, and whether it used a primary-key anchor.
Queries that exceed your tenant’s cost budget are rejected with a suggestion to narrow the scope.

​
Connect a client

The server speaks Streamable HTTP and uses OAuth 2.1 with Dynamic Client Registration. Any MCP client that supports remote HTTP MCP servers can connect. The first time you connect, your client pops a browser window for sign-in. After that, refresh tokens are handled automatically by your client.
You do not need to provision API keys, copy bearer tokens, or paste anything from the Origin console. The OAuth flow handles everything.
Add this to your claude_desktop_config.json (Settings → Developer → Edit Config):
Restart Claude Desktop. The first time you reference Origin in a chat, you’ll be prompted to sign in.

​
Sample prompts

Once connected, you can ask questions in plain English. Examples to try:
  • “What models are my Claude Code agents calling this week, and on which endpoints?”
  • “Show me total token spend by department over the last 30 days, sorted highest first.”
  • “Which prompts are failing most often? Group by model and tool name.”
  • “Find the ten longest agent sessions yesterday and show me which tools each one called.”
  • “Compare prompt latency p50 and p95 across our models for the past two weeks.”
  • “Which endpoints had AI activity for the first time this week?”
  • “How many Claude Code sessions did each engineer run this week?”
  • “Show me MCP server connections by endpoint — which agents are connecting to which servers?”
For the best results, give the LLM enough specificity to compose a query: include a time range, a grouping dimension, and what you want to measure. If the LLM asks clarifying questions, answer them — it’s usually trying to disambiguate between similar fields.

​
Troubleshooting

Most often this means the client hasn’t finished discovery. Fully quit and reopen the client. If the tools still don’t show up, check the client’s MCP logs for connection errors and confirm the endpoint URL is exactly https://mcp.prod.originhq.com/mcp with no trailing path.
Your client needs an OAuth-capable browser flow. Make sure you’re on a recent version — older Claude Desktop, Cursor, and Cline builds do not support OAuth 2.1 Dynamic Client Registration. Update and retry.
Your token may have expired and your client may not be refreshing it. Disconnect and reconnect the server in your client’s settings. If the problem persists, your Origin account may not have access to the tenant you’re trying to query — confirm in the Origin console that you can see the data there first.
The LLM composed a query the server couldn’t validate — usually a field referenced on the wrong entity, or a filter value of the wrong type. These come back as visible errors so the LLM can self-correct on the next turn. If it loops, try restating your question with the field names you saw in the schema, or ask the LLM to call analytics_describe_schema again before retrying.
You asked for something that would scan more rows than your tenant’s per-query budget allows. The error message includes a suggestion — usually: narrow the time range, add a filter, or sample the result. Time-bound your question more tightly and try again.
The LLM may have applied a filter that doesn’t match anything. Ask it to show the query it ran — most clients will print the DSL — and check the filters. Common culprits: a hostname filter with the wrong case, or a time range that excludes the data you care about.
Analytics queries scale with the data they scan. If a question consistently takes more than 10–20 seconds, try narrowing the time range or asking for an aggregate (counts, top-N) instead of raw rows.

​
Versioning and protocol

Discovery metadata is published at https://mcp.prod.originhq.com/.well-known/oauth-protected-resource for clients that need it. You should not need to configure this manually.

​
Data, privacy, and audit

Every tool call is recorded to Origin’s audit log with the calling user, the tool name, a hash and snapshot of the arguments, the row count, bytes scanned, elapsed time, and outcome. Audit records are scoped per tenant and retained for 30 days.
Queries run with your user identity — you can only see what your Origin role allows you to see in the console. See Access Control.

​
Get support

When reporting an issue, include your MCP client and version, the prompt you ran, and any error message the client returned. If the server returned a structured error, the message is usually enough to pinpoint the problem.