Set up Measure Agent

Use this guide to set up Measure Agent on a self-hosted instance, so you can debug your app from your coding agent (over MCP) or from Slack.

The agent runs as a separate agent service and sends prompts to a language model through OpenRouter. You provide an OpenRouter API key and choose which models the agent uses.

Get an OpenRouter API key

1. Create an account at openrouter.ai.
2. Add credits to your account on the Credits page. OpenRouter bills the agent's usage per token, so the account needs a positive balance for the agent to answer questions.
3. Create an API key on the Keys page and copy it. You will set it as LLM_AGENT_KEY.
4. Optionally, create a second key for the Ask AI chat on the docs pages. You will set it as LLM_DOCS_CHAT_KEY. Give this key a credit limit, so traffic from the public docs pages cannot spend the agent's credits. Skip it to keep the docs chat off.

Choose the models

The agent uses two models:

• Small model (LLM_AGENT_MODEL_SMALL) handles light work: summarizing long conversations and working out which app a Slack question is about. It does not call tools, so any capable chat model works. Pick a fast, inexpensive one.
• Medium model (LLM_AGENT_MODEL_MEDIUM) answers the questions by calling Measure's tools, so it must support tool (function) calling. Pick a stronger model, since this is where most of the answer quality comes from.

Both models should have a context window of at least 64k tokens so they can hold a long conversation comfortably.

You can use the same model for both, as long as it supports tool calling. Browse the available models at openrouter.ai/models, confirm the medium model lists tool (function) calling support, and use the model's API id, for example deepseek/deepseek-v4-pro.

Configure for a new installation

During installation, the configuration wizard prompts for the agent settings. At the agent LLM credentials and models prompts, paste your API key and the small and medium model ids. The wizard writes AGENT_ENABLED=false, so once installation finishes, turn the agent on as described in Turn the agent on.

The wizard also asks for the Measure Agent service URL, for example https://measure-agent.yourcompany.com. This is the public address coding agents use to reach the MCP endpoint. See Make the agent reachable over MCP.

Configure for an existing installation

If your instance is already running and you want to enable or change the agent, update the environment variables manually.

1. Agent Credentials. Open the self-host/.env file & add the following environment variables. Set the credentials and models as obtained from OpenRouter, and set AGENT_ENABLED=true to turn the agent on.

   AGENT_ENABLED=true
   LLM_AGENT_KEY=your-openrouter-api-key            # change this
   LLM_AGENT_MODEL_SMALL=deepseek/deepseek-v4-pro   # change this
   LLM_AGENT_MODEL_MEDIUM=deepseek/deepseek-v4-pro  # change this
   LLM_DOCS_CHAT_KEY=your-docs-chat-api-key         # optional, turns on the docs Ask AI chat

2. Shutdown. Run the following command to shutdown all services.

   sudo docker compose -f compose.yml -f compose.prod.yml --profile migrate down

3. Startup. Finally, run the install.sh script for the configuration to take effect.

   sudo ./install.sh

Turn the agent on

The agent is off by default. It answers only when AGENT_ENABLED is set to true in self-host/.env; with any other value, or with the variable missing, Slack questions and the MCP ask_question tool receive a short notice that the agent is unavailable instead of an answer. The other MCP tools read data directly without the agent, so they keep working.

To turn the agent on, set the variable and restart the services as described above.

AGENT_ENABLED=true

Set it back to false and restart to take the agent out of service without removing its configuration.

Make the agent reachable over MCP

To use the agent from a coding agent over MCP, the agent service (port 8084) needs a public domain. This uses the same reverse-proxy and DNS setup as the dashboard, API and ingest services in the Self-Hosting Guide.

Add a reverse proxy entry for the agent domain. For Caddy, add the following to your ~/Caddyfile.

measure-agent.yourcompany.com {
	reverse_proxy http://localhost:8084
}

Reload the proxy, then add a DNS A record pointing measure-agent.yourcompany.com at your VM's external IP, the same way you did for the other subdomains.

The MCP endpoint is then available at https://measure-agent.yourcompany.com/mcp. See the MCP Server guide for connecting coding agents.

Using the agent from Slack needs no extra domain: Slack delivers events to the API service and the agent consumes them internally. Set that up with the Slack integration guide.