fal Agent (early access) vs Sume Agent Completions and hosted MCP

fal Agent, announced August 12, 2026 and in early access, orchestrates image, video and 3D models and is programmable through an API, a CLI and an MCP server. Sume's agent is reachable the same three ways, with POST /v1/agent/completions, the CLI, and a hosted MCP at https://mcp.sume.com/mcp, but the two differ in how much each documents.

fal's facts are from its launch press release and Sume's from Agent Completions, Agents and the MCP overview, all read on 2026-10-02.

What does the fal Agent announcement say?

The release describes a creative partner that works across top generative media models, picks a model per task, keeps characters and references consistent across generations, and preserves project memory over multi-week iterations. It supports shared workspaces and is available in early access. Customers buy add-on credits usable across the fal platform, and the release does not disclose specific pricing.

Because it is a press release, it gives no endpoint names, request shapes or limits. Anything beyond that, wait for fal's API docs.

What does Sume document for its agent?

An Agent Completion runs the Sume Agent on an ad-hoc prompt with the same runtime as the Agents chat UI: sandbox, tools, MCP bridge and media generation. It is async: the create call returns 202 and a receipt, which you poll at /v1/agent-runs/:run_id/status. The request borrows the messages[] shape, but the response is a run receipt, not choices[]. Streaming and a sync OpenAI-compatible wire are not available yet.

generation_spend_cap_usd has no default; omit it and the request fails. Keys made before the feature shipped lack the agent_completions scopes, and service-account keys cannot create completions.

Which should you pick for a backend?

If you need a documented, callable contract today, Sume's is the one with request fields, scopes, receipts and a cancel route (POST /v1/agent-runs/:run_id/cancel). If you already generate on fal and want its orchestration, fal Agent is the one to try, but its press release is not a spec.

Sume's other two entry points share the same receipt: a saved Format via POST /v1/formats/{handle}/{slug}/runs, and scheduled runs for recurring jobs.

What does Sume not do?

Sume's hosted MCP is the preferred MCP path today; a local CLI MCP surface is not available. Paid MCP tools need an idempotency_key, and with OAuth, paid tools appear only after you opt into mcp:write. Sume Image 1.0 and Video 1.0 stay REST-only. Sume also does not claim project memory across weeks as a documented feature on these pages, so do not expect parity with fal's description. See the fal MCP setup post for the fal side.