Check LLM Usage and Costs From Your Coding Agent

You are three hours into an agent session and want to know one thing: what has this cost so far, and which model is doing the damage? Until now that meant tabbing out to the dashboard. With the LLM Gateway MCP server's new analytics tools, LLM usage and costs are one question away in the same terminal, and the answer comes from the same numbers the dashboard shows.

This guide connects the MCP server to Claude Code, Codex, or Cursor and walks through the questions it can answer.

Step 1: Create an API key

Open the API Keys section of the dashboard and create a key for the project you want to inspect. Analytics are scoped to that project, so use one key per project you care about. A DevPass plan key works too.

Step 2: Connect your client

Claude Code

1claude mcp add --transport http --scope user llmgateway https://api.llmgateway.io/mcp \2  --header "Authorization: Bearer $LLM_GATEWAY_API_KEY"

1claude mcp add --transport http --scope user llmgateway https://api.llmgateway.io/mcp \2  --header "Authorization: Bearer $LLM_GATEWAY_API_KEY"

Codex CLI

1codex mcp add llmgateway --url https://api.llmgateway.io/mcp \2  --bearer-token-env-var LLM_GATEWAY_API_KEY

1codex mcp add llmgateway --url https://api.llmgateway.io/mcp \2  --bearer-token-env-var LLM_GATEWAY_API_KEY

Cursor (~/.cursor/mcp.json, v0.48.0 or later)

1{2  "mcpServers": {3    "llmgateway": {4      "url": "https://api.llmgateway.io/mcp",5      "headers": {6        "Authorization": "Bearer your-api-key-here"7      }8    }9  }10}

1{2  "mcpServers": {3    "llmgateway": {4      "url": "https://api.llmgateway.io/mcp",5      "headers": {6        "Authorization": "Bearer your-api-key-here"7      }8    }9  }10}

Run /mcp in Claude Code or Codex to confirm the llmgateway server is connected. Any other client that speaks Streamable HTTP MCP works the same way; the MCP docs cover the generic configuration.

Step 3: Ask

Three tools back the analytics, and your assistant picks the right one from a plain question:

get-usage takes an optional from and to (UTC dates, inclusive, default the last 30 days, at most 366 days) and a granularity of day or hour (hourly at most 31 days). get-usage-breakdown takes group_by of provider, model, app, or api_key, sorts by requests, cost, or tokens, and pages with limit and offset:

1{2  "group_by": "model",3  "sort_by": "cost",4  "from": "2026-08-01",5  "to": "2026-08-31",6  "limit": 57}

1{2  "group_by": "model",3  "sort_by": "cost",4  "from": "2026-08-01",5  "to": "2026-08-31",6  "limit": 57}

Step 4: Read the numbers

Every get-usage response carries totals (requests, errors, cache hits, input, output, and total tokens, and costs), a series of only the buckets that had activity, and mostUsedProvider, mostUsedModel, and mostUsedApp by request count. Breakdown rows carry the ID, display name, requests, tokens, and the same cost fields.

These are usage statistics, not invoice totals: they come from hourly aggregates, survive request-retention cleanup, and may lag the newest requests by a few minutes.

Step 5: Attribute usage to apps

App rankings rely on each request's recorded source. Recognized coding clients are attributed automatically, known aliases are merged before ranking, and anything with no recorded source shows up as unknown. If your client does not identify itself, set an x-source header on the MCP connection; generation calls made through MCP forward the client's attribution headers unchanged.

Per-key app statistics start when per-key source aggregation was enabled, so older periods may report coverage.complete: false. That means the ranking covers part of the period, which is different from an unknown source.

What the tools will not do

• Reach outside the connected project. Owners and admins see the project across all of its keys; developers see only the keys they created. No scope parameters exist.
• Return credentials. get-account reports the user, organization, project, role, scope, and spending limits, and owners and admins also see the credit balance, but never a key.
• Fabricate a zero-usage report. An unreachable backend produces a tool error.
• Stop at a spending limit. Analytics stay available when generation tools are blocked, and calling them is free.

Getting started

• Create an API key and connect your client with the snippet above
• MCP docs list every tool, parameter, and client configuration
• Compare usage across periods in the dashboard when you need the chart instead of the number