Origin MCP Server: Query Observability Data from AI Tools

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 will translate them into structured queries against your Origin tenant.

What the MCP server gives you

Origin’s MCP server exposes read-only access to your analytics graph — the same data that powers the Origin dashboard. 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 MCP 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. The LLM fills in the parameters for you based on your natural-language question.

Parameter	Description
`entity`	The root entity to query — e.g. `endpoints`, `ai_prompts`, `sessions`
`joins`	How to traverse into related entities
`filters`	Predicates applied before aggregation
`select`	Dimension fields to return, optionally bucketed by time
`metrics`	Aggregations to compute: counts, sums, percentiles, computed metrics
`order_by`	Sort order for results
`limit`	Maximum number of rows to return
`offset`	Row offset for paging
`time_range_days`	Lookback window in days; defaults to 30
`sample`	Distribution-preserving sampling for expensive queries

The response includes the result rows, a truncated flag indicating whether rows were cut off, 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. Endpoint: https://mcp.prod.originhq.com/mcp Authentication: OAuth 2.1 via your Origin login (WorkOS AuthKit). The first time you connect, your client opens 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 dashboard.

Choose your client and add the server

Select the tab for your MCP client and follow the instructions.

// Add this to claude_desktop_config.json
// Open it via: Settings → Developer → Edit Config
{
  "mcpServers": {
    "origin": {
      "url": "https://mcp.prod.originhq.com/mcp"
    }
  }
}

claude mcp add --transport http origin https://mcp.prod.originhq.com/mcp

// Settings → MCP → Add new MCP server
{
  "mcpServers": {
    "origin": {
      "url": "https://mcp.prod.originhq.com/mcp"
    }
  }
}

Cline and other clients: Point your client at https://mcp.prod.originhq.com/mcp and follow your client’s instructions for completing the OAuth sign-in flow.

Restart your client

After saving the configuration, fully quit and reopen your client so it picks up the new server entry and runs discovery.

Claude Desktop: Restart the app entirely.
Claude Code: Run claude in your terminal — the server is added to the session immediately after the mcp add command.
Cursor: Restart or reload the MCP settings pane.

The first time you reference Origin data in a conversation, your client opens a browser window for sign-in. Complete the Origin login (WorkOS AuthKit) and return to your client. After the initial authorization, refresh tokens are handled automatically — you won’t need to sign in again unless you disconnect the server or your session is explicitly revoked.

Sample prompts

Once connected, ask questions in plain English. The LLM calls analytics_describe_schema to learn your data model, then uses analytics_query to fetch results. For best results, include a time range, a grouping dimension, and what you want to measure.

“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?”

If the LLM asks clarifying questions, answer them. It is usually trying to disambiguate between similar fields or resolve an ambiguous time range.

Troubleshooting

No Origin tools appear in my client after adding the server Most often the client hasn’t finished discovery. Fully quit and reopen it. If tools still don’t appear, 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. Sign-in window doesn’t open, or the flow is stuck on “authorize” Your client needs an OAuth-capable browser flow. Make sure you are on a recent version — older builds of Claude Desktop, Cursor, and Cline do not support OAuth 2.1 Dynamic Client Registration. Update your client and retry. “Authorization failed” / 401 errors after sign-in 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, confirm in the Origin dashboard that you can see the data there — your Origin account may not have access to the tenant you are trying to query. “Validation error” or “DSL → query” errors in responses 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 errors are returned as visible messages so the LLM can self-correct on the next turn. If it loops, restate your question using field names from the schema, or ask the LLM to “call analytics_describe_schema again before retrying.” “Cost exceeds budget” errors Your query 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 enable sampling. Time-bound your question more tightly and try again. Empty results when you expected data 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 review the filters. Common culprits: a hostname filter with the wrong case, or a time range that excludes the data you care about. Slow responses 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

Property	Value
MCP protocol version	`2025-06-18` (with fallback to `2025-03-26` and `2024-11-05`)
Transport	Streamable HTTP (`POST /mcp`)
Auth	OAuth 2.1 with Dynamic Client Registration (RFC 7591) and Protected Resource Metadata (RFC 9728)
Request body limit	256 KiB

Discovery metadata is published at https://mcp.prod.originhq.com/.well-known/oauth-protected-resource for clients that need it. You do 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 dashboard.

Get support

In-app: Open the support panel in the Origin dashboard.
Email: support@originhq.com
Docs: support.originhq.com

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.

Dashboard

AI Assistant

Configuration

Origin MCP Server: Query Observability Data from AI Tools

What the MCP server gives you

Tools exposed

`analytics_describe_schema`

`analytics_query`

Connect a client

Sample prompts

Troubleshooting

Versioning and protocol

Data, privacy, and audit

Get support

​What the MCP server gives you

​Tools exposed

​analytics_describe_schema

​analytics_query

​Connect a client

​Sample prompts

​Troubleshooting

​Versioning and protocol

​Data, privacy, and audit

​Get support

What the MCP server gives you

Tools exposed

`analytics_describe_schema`

`analytics_query`

Connect a client

Sample prompts

Troubleshooting

Versioning and protocol

Data, privacy, and audit

Get support