MCP Server

Claude Code

Add the remote server from the terminal.

claude mcp add --transport http agentcost https://agentcost.tech/api/mcp

Claude Desktop, Cursor, Windsurf and other config-file clients

Add an entry under mcpServers in the client's config file.

{
  "mcpServers": {
    "agentcost": {
      "type": "http",
      "url": "https://agentcost.tech/api/mcp"
    }
  }
}

Anything else that speaks MCP

Point any MCP client at the endpoint over Streamable HTTP. No install step, no OAuth flow.

https://agentcost.tech/api/mcp

With your project API key

The account tools read your own project, so they need its API key as a bearer token. The pricing tools keep working without one.

claude mcp add --transport http agentcost https://agentcost.tech/api/mcp \
  --header "Authorization: Bearer sk_your_project_api_key"

# config-file clients
{
  "mcpServers": {
    "agentcost": {
      "type": "http",
      "url": "https://agentcost.tech/api/mcp",
      "headers": { "Authorization": "Bearer sk_your_project_api_key" }
    }
  }
}

list_models

Search the public AgentCost catalogue of LLM models and their per-1,000-token rates. Filter by provider and/or a substring of the model name. Use this to compare model costs or find a cheaper alternative. Returns rates in USD; no credentials required.

provider

Provider slug to filter by, e.g. openai, anthropic, google, aws, azure.

query

Case-insensitive substring to match against the model name, e.g. 'sonnet' or 'gpt-4'.

limit

Maximum models to return. Results are sorted cheapest-input-first.

get_model_pricing

Look up the per-1,000-token rates for a single named model. Resolves the name exactly, then case-insensitively, then by provider-prefixed suffix (so 'gpt-4o' finds 'azure/gpt-4o'). Use this when you already know which model you are pricing.

model*

Model name as the provider spells it, e.g. gpt-4o or claude-sonnet-4-6.

estimate_cost

Work out the USD cost of one or more LLM calls before making them, from the current catalogue. Give a model and the token counts involved. Use `calls` to price a whole job rather than a single call. Cached input tokens bill at the provider's cached rate when one is published and at the full input rate when it is not; the result reports which happened.

model*

Model name, e.g. gpt-4o.

input_tokens*

Prompt tokens per call.

output_tokens*

Completion tokens per call.

cached_input_tokens

Prompt tokens served from the provider's cache, per call.

cache_write_tokens

Tokens written into the prompt cache, per call.

calls

How many identical calls to price. Use this to cost a whole job.

list_model_deprecations

List models with an upstream-announced retirement date, soonest first. Use this to check whether a model a user depends on is being retired, or to avoid recommending one that is.

provider

Optional provider slug to filter by.

limit

get_spend_overview

Total LLM spend, call count, tokens, average cost per call and success rate for the caller's own AgentCost project over a time window. Needs the project API key in the Authorization header.

range

Time window ending now.

get_spend_breakdown

Cost and call volume for the caller's own project, grouped by agent, model, workflow, tool, user or session, most expensive first. Use this to find which agent or model is spending the money. Needs the project API key in the Authorization header.

dimension*

What to group by. user and session come from user_id and session_id in event metadata.

range

Time window ending now.

limit

Maximum groups to return.

get_budget_state

Month-to-date spend against the project's budget: the budget, what is left, utilisation, whether it is exhausted and when the period ends. Use this before starting an expensive job. Needs the project API key in the Authorization header.

get_run_cost

Every model call in one run (trace) of the caller's own project, in execution order, with the cost of each, the run total and how the run ended. The trace id is the one the SDK's workflow() minted or an external system supplied. Needs the project API key in the Authorization header.

trace_id*

The run's trace id.