Let a video agent read before it spends: MCP OAuth read, then write

How do you stop an AI agent from spending money while it plans a video? Connect it to hosted MCP with OAuth and leave the Write toggle off: mcp:read is required and read-only. Turn on mcp:write on the consent page, or use an API key, only when you want it to run paid tools such as avatar creation.

Sume's docs treat reads and paid writes as two different permissions, so a planning agent can look without being able to spend.

The matrix

Hosted MCP lives at https://mcp.sume.com/mcp. It accepts OAuth access tokens or Sume API keys, and one credential cannot stand in for the other.

What a planning pass reads

With mcp:read, a session sees only read-only tools; a call that changes data returns insufficient_scope. The paid hosted tools include generate_image, generate_video, music_create, tts_create, avatars_create and avatar-videos_create. Write and paid calls must send idempotency_key.

There is no mcp:paid scope; write is the gate. The workspace comes from the key or the session, so tools must never ask the user for a workspace id.

A safe default

Start every new agent integration with OAuth and Write off. Let it list catalog models, read job status and usage, and draft a plan with prices. Review the plan, then grant write for the session that executes it.

Log request ids, job ids and high-level status. Do not log API keys, signed URLs, raw private media URLs or large transcripts.

Where this fits

This split suits agents that plan a season of episodes. The planning pass uses read tools only. The execution pass needs write. Before the first paid submit, Sume recommends dry_run=true or generation_admission_preview, and max_spend_usd sets a maximum spend when you pass it. A person reads the plan, then enables write.

• Use OAuth for people-in-the-loop work.
• Use an API key for unattended jobs, with a tight cap.
• Send idempotency_key on every write and paid call.

Reviewing the plan

A useful plan lists each call the agent intends to make, the model or tool, and the expected cost. Ask the agent for that list in structured form, use dry_run to preflight cost, and only then switch on write. If the agent proposes something outside the list, deny the call.