scenario-labs / scenario

Scenario

Overview

Scenario (scenario.com) generates AI images, video, 3D, and audio across 500+ models plus custom training, all through the core loop below.

Setup

Endpoint: https://mcp.scenario.com/mcp (Streamable HTTP). Prefer OAuth: no credentials pass through the conversation. Client config, API-key setup for headless use, and per-client re-authentication: [references/setup.md](references/setup.md). Never ask an agent to collect, encode, or echo a secret. A connection that authenticated once and fails later is re-authenticated per client (setup reference) before anything else is debugged; diagnostics_run names the failing layer (auth, tenant-scope, api-unreachable, or healthy).

The default toolset is wider than the loop below: asset_get, job_get, jobs_list and models_list are called directly too. ?toolsets=full exposes everything. The catalog tools are the ones outside it (collections, tagging, analysis, training, members, keys): scenario_tools_search with the tool name or plain keywords as query (it takes only query and limit) returns the schema and lane, and the matching scenario_tool_execute_read / write / delete runs it with {name, parameters}. Every input goes inside parameters, dry_run and scope ids included (nearly every catalog tool declares scope ids; plan_generation takes only description), though direct tools take scope, and model_run its dry_run, at top level. The executor drops a misplaced key without an error: arguments surfaces as a false scope error, and a top-level dry_run runs the job, billed. The lane is the result's own permission, not what the verb sounds like: asset_download and asset_analyze are both write-class.

Scope first

Resolve scope first, then pass team_id and project_id on every later call. teams_list returns the teams with their projects; projects_list requires a team_id, so it cannot come first. Confirm the pair with the user: a guess writes into someone else's project. A non-interactive run takes the pair from its task instructions; when they name none, stop and list the choices.

The server fills scope in only for read-only tools with one candidate remaining; anything else fails rather than guesses, and the error names which half is wrong (see Errors and recovery). Scope errors are the most common failure here, surfacing mid-session on the first call that drops the pair once a second team or project is in play.

Quick reference

A multi-step request ("product video with voiceover", "concept to 3D") goes to plan_generation (catalog-only, read lane): plain words in description, ordered steps out, each naming a tool and optional model hint; it runs nothing. Single-step: recommend. A stated preference ("always 9:16") goes to catalog memory_set with scope project_user_memory (user_memory across projects): it replaces the whole layer, so memory_get and merge first, and it runs on the delete lane.

Worked example

Generating a stylized game prop image:

1. recommend with the user's own words as prompt when the need is a capability; search with target="models", query="flux", public=true when you have a name. search ranks by keyword and its filters hold no capability key, so a capability-worded query can rank the wrong output type first. For the user's own trained models, omit public on search (search has no private flag); on recommend the flag is include_private_models: true. Re-discover ids each time: availability differs per team.
2. model_schema_get on the pick: exact field names, types, required flags, defaults, and caps such as the prompt's max_length (an overrun is a 400, never a trim). File fields take asset ids even when named ...Url, and cost_impact: true flags what moves the price.
3. If the schema carries runs_as ("lora" or "composition"), never send that model's own id to model_run. Its run_with.required_arguments holds the real call: model_id there is the base model, and its parameters (the loras or modelId wiring) merge into inputs from the same schema. Sending required_arguments alone discards your prompt.
4. Optional: prompt_spark rewrites a thin prompt into an on-model one; pass the discovered id (a LoRA's own, not its base) and the draft prompt. Skip deliberate prompts.
5. model_run with model_id and schema-conformant parameters. If cost matters (the default assumption unless the user says otherwise), price first with dry_run: true, a top-level argument beside model_id: no job is created and the response's creativeUnitsCost is the exact payload's price; a recommend cost quote assumes defaults, and the schema's cost_impact fields move the real number. Then run: assetids come back, or a `jobidforjobswait. Itsstatusisinprogressafter a timed-out wait, or withwait=falsethe backend's word at creation (queued,in-progress,warming-up`); none is a different state.
6. jobs_wait with job_ids=[...] (up to 32); each completed row carries assetIds and cuCost, so no job_get follow-up; on timeout re-call with the returned pending_job_ids as job_ids. Failed jobs are reimbursed, except xAI generations stopped by moderation.
7. asset_display shows the asset inline; its format picks the rendering: display (the default) returns an inline image plus links and viewer data; viewer returns a lighter payload for a host that renders the interactive widget; json and markdown return metadata and links. display does not disable a host's widget. For PNG files, use asset_download with format: "png", one call per asset. asset_download returns a file URL (save with curl -L, it may redirect). format converts images (png, webp, jpg, and gif to keep an animated GIF animated: the png default flattens it to one frame) and meshes (glb, fbx, obj); omit it for video and audio. When curl reports host_not_allowed or CONNECT tunnel failed, response 403 while MCP calls work, check the host's proxy or sandbox egress policy. A CDN 403 alone does not establish the cause; request a fresh download URL before diagnosing it: Sandbox network access names the setting that lifts it, an organization-level allowlist on some hosts that the user may not be able to change. Meanwhile hand over the app_url that asset_display returns, where the user downloads directly and, for a mesh, picks the 3D export format the app offers, none of which asset_download converts to. The signed URL expires: a last resort, never the deliverable.

Local inputs go up with upload_asset: always file_name, content_type, and kind (image, audio, video, 3d), plus exactly one of file_size or data, since the call fails without either. Prefer file_size, the file's exact byte count read from disk, and omit data: the reply carries presigned part URLs and instructions; PUT each part's raw bytes to its URL with no added headers (a checksum header makes the store answer 403), check every PUT returned 200, then upload_asset_complete with the upload_id. Inline base64 data only under 100KB (that path returns the asset directly, with no complete step); a larger file is rejected naming the cap. Scope rides on both; they take no other fields: no parts list, no etags. The server decodes the file at completion, so Upload upl_… failed: Corrupt JPEG data, premature end of data segment, bad Huffman code or libpng read error means the uploaded image could not be decoded: check for a corrupt source, a truncated or re-encoded body, a missing part, or a mismatched file_size. Confirm the local file opens, its content type matches, and its byte count is correct, then start over with a fresh upload_asset and raw PUTs; never re-complete the same upload_id. A phone's HEIC photo uploads as is (content_type: "image/heic"), so do not convert it first; Unhandled image format lists what kind: "image" accepts, so convert a file of any other type locally. A host with no shell cannot PUT parts at all; the user uploads in the web app at app.scenario.com and the agent continues from the asset id.

Filing is part of delivering, not a tidy-up. collection_create takes a name and the scope pair only; asset_ids sent there is ignored without an error, so adding is always a second call, collection_add_assets with collection_id and asset_ids. It takes any number of ids (it batches them by 49 itself) and skips an id already filed, listing it under already_present_asset_ids beside added_asset_ids, so re-sending the whole list after a failed batch is safe: the error says how many went in, and the retry adds the rest. Confirm from that reply, since search filters={"collection_ids": [...]} lags on a fresh write. asset_add_tags is additive, one asset_id per call, and a tag the asset already carries returns 400 Duplicated tag: read tags via assets_get_bulk first and send only the missing ones.

For reusable templates or reference content, follow the [shared asset lifecycle](references/shared-assets.md): resolve existing assets, stage new versions, publish only through a supported operation, and verify public access before recording public IDs. Uploading and filing alone are not publication.

Errors and recovery

Common mistakes

• Pricing a later step with a placeholder id (asset_placeholder): file ids resolve before pricing, so the dry_run returns 404 and quotes nothing. Quote it from recommend, then dry_run once the real asset exists.
• A bare value where the schema says array: true: silently dropped, the run ignoring your reference or LoRA. asset_get on the output echoes what the run consumed (metadata.referenceImages, parentId), the cheapest proof it was not.
• Taking recommend's ranked[0] blindly: read next_step.type first. On ask_user, present the options; the user's pick wins (non-interactive: task instructions, else proceed). On proceed, prefer specialty.model_id, else the first ranked entry its own text does not mark deprecated, reading tradeoff and explanation as well as caveats, since the flag lands in any of them: the ranking is by measured performance, not by lifecycle, so a deprecated member can top it.
• Debugging blind: the diagnose MCP prompt (or diagnostics_run with the scope pair the failure happened under and any mcpt_ ids seen as observed_trace_ids) returns trace ids; usage answers credit questions.
• Guessing a file field's name: 400 Input image is required (or images, referenceImages, startImage, frontImage, video, model, each a different model's name for the same idea) means the payload never carried the field the schema's required list names, so copy that name verbatim; a URL in a file field reads as missing too, since file fields take asset ids only.
• Feeding an image field a video, audio, or 3D asset: 400 Unhandled image format lists the image types it accepts. asset_get reports the asset's mimeType; pull a still with the clip's firstFrame when the model wants an image.
• filters.kind on search with target="models": a 400 "kind" is not a filterable field, since kind describes assets. Narrow models with filters.type (the architecture) or tags, or go through recommend for a capability.
• Filtering models_list by type without privacy: "public": a 400 says so. A private listing narrows by status or modality, or goes through search (private by default).