Agents and LLMs

Vertex API is built to be driven by agents as easily as by people: every route is documented in machine-readable form, every request can be priced before it runs, and responses are compact.

Give the agent the contract

• https://api.vertexmaps.com/llms.txt is a markdown summary written for models: what exists, how to authenticate, how credits work, the field names, and a curl to copy. Paste it into a system prompt or point a tool at it.
• https://api.vertexmaps.com/v1/openapi.json is the full OpenAPI 3.1 document. Load it into any tool that speaks OpenAPI (function-calling wrappers, Postman, code generators). The reference on this site is generated from it.
• https://api.vertexmaps.com/v1/pricing.json publishes the credit weights, so an agent can estimate spend from a plan.

None of these need a key.

Patterns that keep an agent cheap

• Price first. Add dry_run=1 to a request and read credits before spending. A 50-location, 40-variable, 3-model forecast is 600 credits; the dry run costs nothing.
• Ask for less. forecast_hours=12, current= for right now, daily= for a summary. Fewer numbers in the response means fewer tokens in the context.
• Read the headers. x-credits-remaining and x-credits-daily-remaining say how much room is left; 402 says none. Both are cheaper than a loop of retries.
• Use one key per agent. Usage is tracked per key, so GET /v1/usage (free) tells you what each one has spent.

An MCP server, CLI and skills

A hosted MCP server at mcp.vertexmaps.com, a vertex command-line tool with agent setup, and installable skills for Claude Code and Codex are in development. Until they ship, the OpenAPI document plus any generic HTTP tool is enough: the API is plain GET with a bearer header.

Attribution in agent output

If an agent presents Vertex data to end users on the free plan, the same attribution rules apply as for any other client: “Weather by Vertex API”, and the source-specific lines in Attribution.