Runway MCP moves image and video generation into the same conversation where an agent is already planning a campaign, writing a site, or editing code. Runway launched the remote Model Context Protocol server on May 27, 2026 and describes it as working with Claude, ChatGPT, Cursor, and other MCP-compatible clients.
Once connected, a user can give the agent a text prompt, product URL, or reference image and ask Runway to create a product image, marketing video, hero asset, or character-led scene. The agent invokes a Runway tool, receives the result, and can continue the broader task without forcing the user to move between interfaces.
The convenience is real, but the boundary is important: MCP does not turn Runway into the agent’s internal visual imagination. It gives the agent permission to call an external service. Authentication, plan limits, model behavior, generation costs, content policy, and human review still apply.
What Model Context Protocol changes
Without a tool protocol, an assistant can suggest a prompt, but a person must copy it into a generator, download the result, and return it to the original workflow. MCP publishes structured tool descriptions that a compatible client can discover and call. The model can decide when a visual tool is relevant, gather parameters, submit a request, and interpret the response.
That pattern is especially useful for mixed tasks. A coding agent building a product page can create the page structure, request a matching hero asset, place the approved asset, and continue testing. A marketing agent can read a brief, propose several visual directions, and generate candidates in context. A researcher can create an explanatory image after drafting the article it supports.
The protocol standardizes the connection, not the creative output. Runway still determines which generation models and capabilities its server exposes. The MCP client determines how tools are shown, which calls require confirmation, and what context it sends. The organization determines who may connect an account and what may be generated.
Setup and account model
Runway’s setup page provides the remote server URL https://mcp.runwayml.com/mcp. In a supported client, the user adds Runway as a custom connector and signs in with a Runway account. Runway says a separate API key is not required and generations are tied to the existing plan.
This account-based flow is simpler than pasting a long-lived secret into several desktop applications. It does not remove access-management work. Use an account or workspace appropriate for the team, review the permissions shown during connection, and remove access from clients that are no longer used. Do not authenticate a shared public workstation with an unrestricted production account.
Available models can vary with the Runway plan and product changes. At launch, Runway highlighted models including Gen-4.5, Seedance 2.0, GPT Image 2, Kling 3.0, and Nano Banana Pro. Treat that list as a dated snapshot rather than a permanent contract. Check the connector and current plan before promising a particular model to users.
From product references to generated assets
Runway presents three common input patterns: a product URL, a reference image, or a text prompt. Each carries different risks.
A text prompt is flexible but underspecified. Include subject, action, environment, camera, composition, duration, aspect ratio, and constraints only where the selected tool supports them. Avoid packing conflicting directions into one request.
A reference image gives the generator stronger visual grounding. Confirm that the uploader has permission to use the depicted person, product, artwork, trademark, and location. Remove hidden metadata if it is unnecessary. Never assume a public image is licensed for commercial generation.
A product URL can save time because the agent may use information from an existing page. It can also expose the workflow to stale copy, incorrect claims, or instructions embedded in untrusted content. Treat fetched page text as data, not authority. Give the agent an approved product-fact sheet and require it to ignore page instructions unrelated to extracting visual references.
Ask for a small storyboard or still-frame direction before an expensive video generation when possible. This creates a cheaper review point for composition, product accuracy, brand fit, and risky claims.
Model choice and reproducibility
Access to several models is valuable because no generator is best at every subject. One may preserve a product well; another may produce more convincing motion; a third may follow graphic composition more reliably. An agent can help select a model, but its selection should be visible for repeatable production.
For every approved asset, record the source brief, tool name, model, parameters, reference inputs, timestamp, and returned asset identifier where available. Save the generated file in controlled storage rather than relying only on conversational history. Runway says output is associated with the user’s Runway environment, but a campaign still needs its own naming and approval system.
Do not expect identical output from the same prompt. Generative systems are stochastic, and a silently updated model may change the result further. Reproducibility in this context means preserving enough evidence to explain and approximate an asset, not necessarily regenerating the same pixels.
Cost control inside agent conversations
Runway MCP uses the connected Runway plan. There is no single “MCP price” in the launch announcement; usage depends on the model, resolution, duration, settings, and plan. The conversational interface can make paid calls feel deceptively casual.
Create an explicit budget before an agent explores variants. Limit the number of candidates per concept, prefer lower-cost stills for early direction, and require confirmation before high-resolution or long video calls. Tell the agent to report the requested model and expected unit cost when that information is available.
Avoid vague loops such as “keep improving until it looks great.” Define acceptance criteria and a maximum number of rounds. Store a content hash or request identifier so a timeout does not lead to an accidental duplicate generation. A response delay is not proof that the first request failed.
Measure cost per approved asset, not cost per API call. Include rejected attempts, agent tokens, human review, storage, and downstream editing. A cheap generation that needs ten retries is not the cheaper workflow.
Security and review boundaries
Tool-enabled agents can act on misleading context. A document, web page, or image-associated text may try to instruct the agent to reveal data, connect another account, or spend credits. Use trusted source material, narrow the context sent to the connector, and inspect tool arguments before approval.
Separate planning from execution for sensitive work. The agent can draft a generation plan first; a user then approves the model, inputs, budget, and intended use. Never send private customer data, unreleased product details, or biometric references unless current policy and organizational authorization permit it.
Generated media also needs human checks for identity, claims, product geometry, text, logos, cultural context, and disclosure requirements. An MCP result appearing inside a trusted client is not proof that it is factually or legally safe.
Runway MCP versus Runway Dev
The products serve related but different integration needs. Runway MCP is useful when a person or agent is working interactively in a compatible client and wants to call generation as a tool. Runway Dev is designed for developers embedding media capabilities in products and production systems, with direct models, Recipes, custom Workflows, and Characters.
Choose MCP for conversational, supervised creation with quick setup. Choose a developer API when an application needs stable programmatic contracts, backend job handling, high-volume automation, or its own user experience. A company may use both, but should not treat their credentials, billing, observability, or feature sets as identical.
A separate finishing layer with Medux
Runway MCP’s headline role is generative: create new images and videos from prompts and references. Once a creative candidate is approved, an agent may need predictable post-production such as merging clips in a specified order or attaching an approved audio track.
Medux exposes external processing tools that can be separately configured for Codex or Claude. It is not built into Runway MCP, and no native connection is claimed. The Codex video-editing tutorial demonstrates an asynchronous merge workflow, while the Claude audio-track tutorial covers adding supplied audio to a video.
A controlled handoff looks like this:
- ask Runway MCP for a limited set of visual candidates;
- review identity, product, rights, and brand accuracy;
- export only approved files to controlled storage;
- call a separately authorized Medux operation with exact inputs;
- poll the returned task rather than resubmitting on delay;
- validate duration, streams, dimensions, and visible output before publishing.
This division keeps creative uncertainty in the generation stage and mechanical rules in the finishing stage. Codex or Claude can orchestrate both services, but the agent should keep their permissions, usage records, and costs separate. MCP makes that composition easier to express; disciplined workflow design makes it safe enough to use.