Tavus MCP server and CLI vs Sume hosted MCP: auth and spend caps

Both vendors let an editor agent drive the API. Tavus offers an MCP server at https://mcp.tavus.io/mcp and a tavus CLI; Sume offers a hosted MCP at https://mcp.sume.com/mcp. The difference buyers feel is spend control: Sume's MCP requires an idempotency_key on paid calls and offers dry_run and max_spend_usd to preview and cap cost.

Tavus's pages describe building PALs, defining tools and testing integrations, which is a build-time workflow. Sume's MCP is for generating media as well as managing it.

What Tavus documents

Tavus says its MCP server connects Codex, Claude Code, Cursor or any MCP client over HTTPS, with browser-based OAuth through PAL Maker and per-user keys forwarded as an x-api-key header. The CLI authenticates with tavus auth login, which stores environment-scoped keys in the OS keychain, or with TAVUS_API_KEY for automation, and both surfaces call the same API. The default environment is production; TAVUS_ENV=TEST targets the test database (Tavus docs: Agent Tools).

What Sume's hosted MCP does

Sume's hosted MCP accepts OAuth access tokens or Sume API keys. Under OAuth, mcp:read is required and read-only; the user can toggle Write at consent to also grant mcp:write. There is no mcp:paid scope: paid submits go through wallet and admission. With an API key you get the full hosted tool set (MCP OAuth and API keys).

On write and paid tools idempotency_key is required as a dedupe key, not as approval. dry_run=true previews admission and cost without submitting, and max_spend_usd is enforced only when you pass it (MCP tools and gates).

Which to connect

For building a live agent, connect Tavus. For generating clips from an editor while keeping a spend ceiling, connect Sume, start with read-only OAuth, and give an agent a max_spend_usd on every paid call.

• An MCP OAuth token is not an API key. Do not mint a key to work around a missing scope.
• Rotate any API key that appears in logs or chat.

A safe starting setup

Start the Sume connection with OAuth and read-only scope so the agent can list models and read results but cannot spend. Add write consent only when a task needs it, and instruct the agent to call with dry_run=true first and show the cost.

For unattended automation use an API key held in an environment variable, a max_spend_usd on every paid call, and a stable idempotency_key per intended job. A retry with the same key does not create a second job.

What to test first

Connect each server from your editor and run one read-only call. On Sume, list the available tools and ask for the schema of one generation tool before calling it. Then run that tool with dry_run=true and read the cost it returns.

Only then allow a paid call, with a max_spend_usd that you are comfortable losing if the result is wrong.