Runway Agent Guide: Skills, Workflows, MCP and Credit Costs

Runway Agent is the chat side of Runway: you describe a video, it drafts a plan, picks a model for each shot, generates the clips and lays them on a built-in timeline. It ships with every Runway plan, yet every generation it starts spends plan credits, and a stronger language model adds a charge per message on top: 24 credits with Opus 5.5, 40 with GPT-6 Astra. Our verdict is simple. Agent is the quickest way to a multi-shot first cut, provided you switch on approval before it generates. Left on its default, it spends first and explains afterwards.

What Agent does between your prompt and the timeline

Runway released Agent in May 2026 and followed with Agent 2.0 a month later. It can generate images, video from text or from an image, single shots or multi-shot sequences, plus audio, instrumental music and scripted voiceover. Runway says Claude powers the reasoning that plans the video and chooses a model for each shot. That choice spans Runway's own Gen-4.5 and outside models such as Seedance 2.0, Kling and Veo, all billed from one credit balance. On Enterprise, an admin can switch individual models off for the workspace.

Clip length works differently from the Tool screens. Agent has no hard duration limit: when a video runs longer than the chosen model allows, it generates separate clips and joins them in the editor ^[1]. That matters because Gen-4.5, Runway's current flagship, produces 2 to 10 second clips at 720p and has no sound of its own in the web app. Agent covers the gap by adding voiceover, effects and music with Runway's audio tools, or by picking a model that generates audio natively. Output lands at 480p, 720p or 1080p depending on the model and quality setting.

You can feed it more than a prompt. Agent reads PDF, CSV, JSON, TXT, Markdown, SRT and VTT documents, with PDFs capped at 20MB, plus OTF and TTF font files. It does not take Word, Google Docs or Excel files directly, and it cannot pull images out of a PDF. In the composer, typing @ attaches References and typing / calls a Skill. Voice input exists only in the iPhone and Android apps. The web app is keyboard only.

Three settings that decide the bill

Most surprise charges trace back to defaults nobody changed. Agent ships set to generate automatically, which means it starts spending the moment it decides on a plan. The alternative mode, Ask before generating media, pauses each step to show the model, the prompt and an estimated credit cost, and nothing runs until you approve. Runway's own troubleshooting article points to that mode, together with optimizing for cost, when results drift from the prompt. Flip it on first.

The optimization setting steers which model Agent reaches for. Pick Cost while drafting and Quality for the final pass. Custom instructions let you state your own preference, such as one model for product shots and another for talking scenes, but Runway calls these guidance rather than strict rules. Agent also generates only with models your plan includes, so a preference for a model above your tier is simply ignored. Check the plan list first.

The language model is the quiet line item. Sonnet 5.5 is the default and the faster option; Runway does not publish what one of its messages costs. Opus 5.5 charges 24 credits a message and GPT-6 Astra 40, once per message and on top of whatever the message generates. By our arithmetic, at the top-up rate of 1,000 credits for $10, that is $0.24 or $0.40 a message, next to $0.60 for one 5-second Gen-4.5 clip. App Store reviewers complain about exactly this message charge.

Credits return only when a generation fails with an error; a clip you dislike keeps its cost, so your real price per usable clip runs higher than the rate card. Runway's own guidance is to judge spend per usable output, and that is the number to track while you learn which prompts and settings work. Our Runway pricing guide covers plans, resets and top-ups.

Teaching Agent your routine with Skills, Workflows and Brand Kits

Skills arrived in July 2026. The built-in set covers Ad Campaign, Commercial, Mood Board, Motion Graphic, Resize Image, Social Thumbnail, UGC Video and Workflow, each called with a slash. The more useful trick is the custom Skill: once a chat produces something you like, Agent can save it as a reusable Skill and re-run it on a new product or input. Instructions run up to 5,000 characters. A custom Skill starts private and can be shared with the whole workspace.

Since July 2026 Agent can also build, run and find Workflows from a plain-language description. It proposes the nodes and settings for your approval before creating anything, and building costs no credits. Running one needs Standard or higher, and in automatic mode the run starts at once. Agent runs Workflows others share with you but copies, rather than edits, one you do not own. The split is clean: a Skill is saved instructions inside a chat, a Workflow a node pipeline saved as its own workspace asset.

Brand Kits hold reusable reference images, videos and font files, and Agent applies a kit's section descriptions when it generates. Agent can create or edit a kit from chat content or from an uploaded PDF, though it reads only the PDF's text. It cannot build a kit from a website address. The pricing page lists Brand Kits from Pro: one kit, three on Max and up to 100 on Team. Runway Ads keeps its own Brand Kits, separate from those in the main app.

Agent on a shared workspace

The Team plan, at $69 per seat a month or $55 per seat a month billed yearly for 2 to 9 seats, adds the collaboration pieces. Agent Connectors import files from Figma, Dropbox and Notion, and they are Team only: not on Free, Standard, Pro or Max, and not currently on Enterprise ^[2]. Each person connects separately per workspace, and each imported file is capped at 50MB. Team also holds 20 shared Agent Skills, so one person's winning routine becomes everyone's starting point.

Comments are the second lever, and the riskier one. An Editor or Admin can tag Agent in a comment on any asset; Agent does the work and posts the result back into the thread as a new asset. That request spends credits like any Agent generation and starts without asking for approval, whatever mode your chat uses. Viewers cannot tag Agent, which is one way to limit who spends the shared pool. Agent inside projects, by contrast, is an Enterprise feature.

Runway Ads, on every paid plan and on desktop web only, is built on Agent as well. Connect a Meta, Google or TikTok ad account and Agent drafts static and video ads from what already performs, sends a plan with copy and an estimated credit cost, and charges nothing until you approve. Every ad then goes to a Review page that cannot be skipped. The advertising guide covers that flow.

Runway MCP: generating from ChatGPT or Claude

Runway MCP, launched in May 2026, puts Runway's generation inside other assistants: ChatGPT, Claude, Grok Bot, Cursor, Replit, Muse and any MCP-compatible agent. You add it as a custom connector and sign in with a Runway account; any account works, a free one included, and no API key is needed. The models offered depend on your plan, from Gen-4.5, Seedance 2.5, Kling 3.0 and Veo 3.1 to Nano Banana Pro and GPT Image 2. Results land in your Runway library.

Costs follow the web app, not the API. MCP generations spend web-app credits, retries and regenerations included, and run in Credits Mode only, so every step is charged at the credit rate ^[3]. The connection works over Streamable HTTP only. In ChatGPT, the plugin's default permission lets low-risk tools run, which starts generations without asking; set it to Always ask. Brand Kits are read-only through MCP. A generation an assistant started still counts as used, so the no-refund rule applies unchanged.

Do not confuse it with Runway Dev MCP. That server serves coding agents: it looks up models and prices, configures Model Routers and Characters, and traces failed tasks, while billing runs on Runway Dev credits that stay separate from app credits on self-serve plans. Our Runway API guide explains that side.

Where Agent stops

Agent has hard edges worth knowing before a deadline. It does not handle 3D models or the 3D nodes in Workflows. It delivers MP4 only, so ProRes needs Tool mode or Edit Studio on Max. Grok Imagine 1.5 is not available to it at all. Runway itself describes Agent as generation-first: it cannot run unsupervised on a loose brief, and a person reviews the output at each step. Plan for that review time.

Longer pieces still end up elsewhere. Runway's own help suggests its Timeline Studio for quick assembly and a local video editor, with proper color and audio tools, for longer films. On data, Runway's consumer Terms let it train on your inputs and outputs; a no-training promise is an Enterprise term. And enabling an MCP connection, under those Terms, authorizes Runway to pass account data to that outside provider. Read that before connecting a client's assistant.

The bottom line

Runway Agent turns one conversation into a planned, multi-shot draft that draws on Runway's own models and other labs' models from a single credit balance, and Skills and Workflows let a team repeat what worked. Its costs hide in defaults: automatic generation, comment tags that skip approval, premium language models billed per message, and an MCP plugin that may generate without asking. Switch on approval, pick a cheaper optimization when drafting, and judge spend per usable clip, not per prompt.