MCP server that gives AI agents (Claude, Cursor, Codex) access to the full Runware API — image, video, audio, text, and 3D generation, upscaling, background removal, and more.
Server has 8 well-structured tools with explicit schemas, descriptions, and tool annotations. Naming follows verb_noun convention (run, search, upload, list). Descriptions are detailed (100-250 chars) and include usage guidance. However, several tools lack output schema documentation, parameter descriptions are sometimes generic, and error handling lacks recovery guidance. The 'run' tool's additionalProperties design is pragmatic but reduces schema clarity for LLMs. No parameter validation constraints (min/max, patterns) are visible in schemas.
Retrieve Runware account information including balance and usage.
Retrieve the original request and response for a previously executed task. Useful for recovering results or auditing past generations.
List Runware's official, curated model integrations. Returns each model's name, AIR identifier, headline (one-line description), capabilities, and pricing — a human-readable catalog. Call this FIRST whenever the user names or asks about a model that could be first-party (e.g. "FLUX 2 dev", "SDXL", "Veo 3", "Gemma", "Wan 2.5", "Z-Image"), and also for open-ended "what models are available?" questions. Match the user's named model against the returned names. Only fall through to model_search if no curated entry matches.
Store or delete media (images, video, audio, 3D models) in your Runware account. Upload returns a media UUID you can reuse as input (seedImage, referenceImages, etc.); delete removes previously stored media by its media UUID.
Get the parameter schema for a specific model. Returns the JSON Schema describing all accepted parameters, their types, defaults, and constraints. ALWAYS call this before calling run() with a model you haven't used before, so you know what parameters to pass.
Output schemas not documented. Tools return results but LLMs cannot see expected response structure (fields, types, pagination). Breaks downstream tool chaining and forces LLMs to guess field names.
'run' tool uses additionalProperties with union type ['string','number','integer','boolean','array','object'] instead of explicit parameter definitions. LLMs cannot discover valid parameters without calling model_schema first, increasing latency and errors.
Error handling returns raw Runware SDK errors (isRunwareError) with code/message/parameter/taskType fields, but no recovery guidance. LLMs receive 'Error: invalid_model' with no hint to call list_models or model_search.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 63 | 2026-07-28+ | v2 |
Search Runware's Civitai mirror and community-uploaded models — third-party fine-tunes, user uploads, style LoRAs, custom checkpoints. ONLY use this AFTER list_models has been checked and the user's named model is not in the curated catalog, OR when the user explicitly asks for a Civitai or community model. Do NOT use this for first-party models like FLUX, SDXL, Veo, Imagen, Gemma, Wan, Z-Image — those live in list_models.
Upload a custom AI model to Runware (checkpoint, LoRA, VAE, embeddings, etc.). Returns the AIR identifier once the upload completes.
Run an AI inference task on Runware. Supports image generation, video generation, audio generation, 3D generation, upscaling, background removal, captioning, and more. Pass a model AIR identifier and task-specific parameters. Example: { "model": "runware:400@1", "positivePrompt": "a cat", "width": 1024, "height": 1024 }
Parameter descriptions lack constraints. 'search' in model_search has no length limit or format guidance. 'media' in media_storage accepts 'URL, data URI, or base64' but no validation rules visible. LLMs may pass invalid formats.
'account' tool description is vague ('Retrieve Runware account information'). Does not explain when to call it vs other tools, what 'includeCost' controls, or what fields are returned.