MIA MCP server

Connect

Add the server URL to your assistant. The first call opens MIA's sign-in in your browser: sign in with your MIA account and approve the connection. There is no API key to paste, store or rotate.

https://mcp.mia.tech

claude mcp add --transport http mia https://mcp.mia.tech

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

{
  "mcpServers": {
    "mia": { "url": "https://mcp.mia.tech" }
  }
}

Any client that supports remote MCP servers over Streamable HTTP with OAuth works the same way. Clients set up with the older https://mcp.mia.tech/mcp keep working. In Claude Code, run /mcp or start a new session after adding it so the tools load.

How an agent works

Four stateless tools, read in order: Explore → Describe → Run, with mcp_status on demand. The server remembers nothing between calls, so an agent may start at any tool. Every answer except run_query carries next_steps, a list of suggested calls.

1. explore_mcp to see which entities your account holds.
2. describe_entity("MediaSpot") to get its resolvers, schema and examples.
3. run_query with a query built from that schema:

query {
  mediaSpots(first: 3) {
    edges { node { id cost } }
    totalCount
  }
}

Entities come in four families: own (your campaigns, traffic and results), competitive (the market), social and news. The catalog lists only what your account can query.

Tools

All four are read-only.

explore_mcp(family="all", search=None)

{ schema_version, user, account, credits, guide, shared_types, entities, field_hits?, warnings?, next_steps }

describe_entity(name)

{ entity, family, scope, resolvers, key_fields, schema, examples, alternatives, next_steps }

run_query(query, variables=None)

{ data, errors?, extensions: { credits: { spent, remaining, cycle_end }, warnings? } }

mcp_status()

{ reachable, latency_ms, credits: { used, limit, remaining, scope, cycle_end, limiting_factor, limit_reached, daily }, manage_connections_url, next_steps }

Prompts

Tools are called by the model; prompts are started by you. Clients show them as slash commands or in the connector menu. In Claude Code, with the server added as mia, usage is /mcp__mia__usage.

usage

This cycle's credit consumption: used, limit, what is left, which ceiling applies, and the daily series.

build_query(question)

Turns a plain-language question into a working query, using the catalog instead of guessed field names.

period_performance(days=7, brand?)

The last N days against the N before, over the metrics your account actually has.

explore_entity(entity)

What one entity holds, which fields can be filtered, and one small query that works.

Errors and warnings

run_query reports problems in the standard GraphQL shape, each with a code in errors[].extensions.code. Connection and limit failures use the same shape.

{ "errors": [{ "message": "…", "path": ["mediaSpots"], "extensions": { "code": "unknown_field" } }] }

unknown_resolver, unknown_field, invalid_filter, orderby_not_selected, invalid_groupby

The query does not match the schema. Call describe_entity, fix it and run again.

invalid_argument

A wrong or missing argument value: currency, id, type or enum.

invalid_query

Syntax error, a mutation or subscription, several operations, or another document error.

limit_exceeded

The query is too long, too deep or uses too many aliases. Narrow it.

credits_exhausted, user_credit_limit_reached

The account's credits, or your own cap, are used up. Call mcp_status.

auth_invalid

The connection is no longer valid. Reconnect and sign in again.

rate_limited

Too many queries in a minute. Carries retry_after_seconds. This is not a credit limit.

backend_unreachable, backend_error

MIA could not be reached or answered unreadably. Retry, or call mcp_status.

Warnings arrive in extensions.warnings as {code, text, path}, next to a successful answer:

own_brands_only

Your own data holds your brands only. The warning names the competitive entity that holds the others.

default_currency_usd

A money field was selected without a currency, so results are in USD.

credits_low

10 credits or fewer remain after the query.

Credits and limits

• Queries spend credits from your account's pool, shared by every member. An account manager can also set a cap per member. Every run_query response reports what it spent and what is left.
• Credits renew at the end of each cycle (cycle_end). mcp_status shows the full breakdown.
• A query is at most 20,000 characters, 10 levels deep, 15 aliases and 2,000 tokens.
• A short per-minute burst limit may apply; it answers rate_limited with the seconds to wait.

Access and revocation

• Read-only. The server can run queries and nothing else. It has no way to change, delete or create data.
• Your permissions. Every query runs as the signed-in user, so it sees exactly what you see in the MIA platform.
• Your token stays with MIA. The assistant only holds a reference token; your MIA session never leaves the server.
• Revoke any time in the MIA platform under Settings → MCP connections. It takes effect on the next call.

← MIA MCP overview