serve (MCP) parity: expose video extend / scene / chain, and a models-catalog tool #880

Activity

1. ffroliva commented on Sep 20, 2026

   

   ffroliva

   on Sep 20, 2026

   Owner

   More actions

   This is a good issue — it names the contract correctly and it names three gaps that are genuinely different from each other. Taking them in reverse order of difficulty, because your framing has the sizing inverted on two of them.

   Gap 3 — models catalog: agreed, and it is the smallest thing in this issue

   You are right, and the current exemption is wrong. It reads:

   "models": "informational; models are enumerated in the generate tools' descriptions"

   Prose in a tool description is not a catalog. It has no ref_cap, no max_duration, no alias list, and nothing machine-readable — which is exactly why you ship a dated snapshot. That exemption should never have survived an option-level look.

   The good news is it is nearly free. cli_models.build_catalog() is already pure, Click-free and does no I/O — it derives everything from the Model / VideoModel enums and the alias maps, so it cannot drift from what the generation commands accept. The tool is a wrapper over it.

   And your instinct about projectInitialData is right: a catalog tool must not touch the labs tRPC listing. Beyond the 401s you are hitting, project.createProject on that surface now answers 404 "Flow RPCs have been deprecated and disabled". A tool reading the static table stays correct on both hosts and spends nothing.

   Filed as #883. Yours if you want it — see below.

   Gap 1 — extend: agreed in principle, but it is not mechanical, and it is blocked

   Two things in front of it.

   It is not an adapter. The exemption's reasons are still live: an extend chain runs for minutes and is not wired to FlowWorker, so a direct-execution MCP tool blows the client's tool-call timeout — that is #481, and it is the reason gflow video chain sits in the same bucket. The port is "enqueue through the worker like gflow_generate_video does", not "wrap the CLI args". It also spends credits behind a confirmation prompt that has no MCP equivalent, so it needs a GFLOW_MCP_NO_SPEND decision before it ships.

   #882 is in the way. There is an open PR rewiring extend onto flow.google.com via batchexecute fZytfe (it is good work — live-verified with a boundary-frame comparison). An MCP twin written now conflicts with it. That PR has an open question about a bypassed project guard on the migrated arm; once it lands, the twin is unblocked.

   Filed as #884, blocked on #882 and #481.

   Gap 2 — split, because scene and chain are not the same ask

   gflow scene create is a fair port. It composes clips the user already paid for; it is free and server-side. That is the same reasoning that made data download non-exempt in #865 — an agent whose pipeline needs assembly is stranded exactly as a CLI user would be. Filed as #885.

   video chain and movie run stay exempt for now. Both are minutes-long, credit-spending, and not worker-enqueued — the #481 shape again. They come back the moment the worker-enqueue work lands, and #884 is the one that pulls that thread.

   On the offer — yes, please

   You have landed #793 before, so this is not a cold start.

   #883 (models catalog) is yours if you want it. Suggested shape, so we do not discover a disagreement at review time:

   gflow_list_models() -> dict
   # returns build_catalog() verbatim:
   #   {"image": {"models": [{name, aliases, ref_cap, default}], "aspects": [{ratio, wire}]},
   #    "video": {"models": [{name, aliases, ref_cap, max_duration}], "aspects": [...], "tiers": [...]}}

   No arguments, no profile, no browser, no spend. Register it beside gflow_list_tools in src/gflow_cli/mcp/tools.py, move "models" from _MCP_EXEMPT to CLI_TO_MCP in tests/mcp/test_cli_parity.py, and add the tool to docs/MCP.md.

   One thing the repo will ask of you that is easy to miss: the MCP twin is its own surface, so it needs its own run. A passing unit test is not the gate here — please exercise the tool over a live serve and paste the result. For a read-only, zero-credit tool that is a ~60s exercise, and there is precedent: the credits twin in tests/e2e/test_credits_mcp_e2e.py is 3 assertions and 58 seconds.

   Say the word and #883 is assigned to you. If you would rather I take it, that is fine too — it is small enough that I would rather it exist than wait.

   Thanks for the detail in the report, and for saying plainly what you are not asking for. That paragraph saved a round trip.
2. ffroliva

   added a commit that references this issue on Oct 7, 2026

   feat(upscale): port image and video upscale to flow.google.com and ex…

   ...

   6f13272
3. kittinan

   added a commit that references this issue on Oct 9, 2026

   feat(upscale): port image and video upscale to flow.google.com and ex…

   ...

   afd6e5d
4. ffroliva commented on Oct 19, 2026

   

   ffroliva

   on Oct 19, 2026

   Owner

   More actions

   Closing this as split into three tracked issues:

   • gap 3 (models catalog) → MCP: expose the models catalog as gflow_list_models #883 (gflow_list_models)
   • gap 1 (video extend) → MCP: expose video extend, enqueued through the worker #884 (extend enqueued through the worker; its blocker feat(extend): drive video extends on flow.google.com via batchexecute fZytfe (#639) #882 has since shipped as feat(extend): drive video extend on flow.google.com through Flow's own UI (#639) #991)
   • gap 2 (scene) → MCP: expose scene create/show (not chain — see #884) #885 (scene create/show)

   video chain stays exempt until worker enqueue lands (#481 / #884). Your offer to contribute still stands on any of them, and #883 is the smallest. Thanks for the clear write-up.
5. ffroliva

   closed this as not plannedon Oct 19, 2026