One markdown file. 45 models across image, video, and audio. Plain HTTP from any agent.
Install the skill once. Claude auto-loads it when relevant.
Sign in at app.runflow.io and create a key in Settings → API Keys. New accounts get $10 in credits. Export it as RUNFLOW_API_KEY in your shell.
Install the skill into a skills directory. Use a repo-local directory when you want the whole project to carry the integration guide.
mkdir -p .agents/skills/runflow curl -fsSL https://www.runflow.io/.well-known/agent-skills/runflow/SKILL.md -o .agents/skills/runflow/SKILL.md
Start a Claude session in any project and ask in plain English. The skill kicks in, Claude picks the right Runflow endpoint, runs it.
> remove the background from ./shoe.png and save the cutout
BetterPic ships 35M+ headshots a year on the Runflow API. 99.9% uptime. Multi-cloud GPU pool.
First call
Real prompt. Real model the agent picked. Real HTTP endpoint. Real image returned.
POST /v1/models/google/nano-banana-pro/edit/runs
Inside the skill
The skill points at every endpoint your agent needs: pre-built workflows for the common tasks, direct model access for everything else, custom ComfyUI graphs for anything you bring yourself.
Nano Banana Pro, FLUX Kontext, FLUX 2, GPT Image 2, Ideogram v3, Wan 2.7, Veo 3.1, Kling v3 Pro, Seedance 2.0. 50 models active. Up to 4K resolution. Any aspect ratio. Any duration up to your model's limit. Your agent picks the right one.
POST /v1/models/google/nano-banana-pro/runs
Background removal, outpainting, object removal, AI headshots, product isolation, skin retouch, tag removal. Each Solution is a multi-step pipeline behind a single endpoint. Your agent follows that endpoint's response, polling, or callback contract.
POST /v1/models/runflow/background-removal/runs
Fan-out is a documented pattern in the skill. Your agent fires the same input at nano-banana-pro, gpt-image-2, flux-pro/kontext, and ideogram/v3 in parallel, returns all four, and lets you choose. No new tool. Just N parallel POSTs.
Drop a graph in the dashboard. We deploy it as POST /v1/comfyui-workflows/{owner}/{slug}/runs. Same auth, same callbacks, same shape as the rest of the API. Your agent invokes it like any other model.
POST /v1/comfyui-workflows/{owner}/{slug}/runs
Batch endpoint takes a list of inputs and a callback URL. Per-item progress tracking. One signed callback when the whole batch finishes. Built for agents that drain queues, not one-offs.
POST /v1/batches
Wan 2.7, Veo 3.1 (Lite + Fast), Kling v3 Pro, Seedance 2.0, HeyGen. Text-to-video, image-to-video. Up to 1080p. Up to 12 seconds. Same auth, same call shape as image models.
POST /v1/models/alibaba/wan/v2.7/text-to-video/runs
Real-world flows
Sample prompts
These are the exact prompts a skill-armed agent handles in one shot. The expected flow is what the agent picks based on the decision rule in SKILL.md.
Generate one finished thing.
“Generate a cinematic wide shot of a neon-lit Tokyo alley at night, anamorphic lens.”
→ The agent picks google/nano-banana-pro, sets the input fields per the model's llms.txt, returns a finished image.
“Remove the background from ./shoe.png and return the cutout as a transparent PNG.”
→ The agent calls runflow/background-removal directly. Solution endpoint, ~1.4s, finished cutout.
Multi-step. Fan-out. Save and ship.
“Take this packshot. Generate 6 lifestyle scenes in different settings, then a 5-second product reveal video for the winner.”
→ The agent runs 6 parallel calls to gpt-image-2/edit with different scene prompts, picks the best, then calls alibaba/wan/v2.7/image-to-video on it.
“Process every product photo in this CSV: bg-removal, outpaint to 16:9, save to S3.”
→ The agent posts a batch to /v1/batches, polls for completion, then writes results back via your callback URL.
Same prompt, different models, side-by-side.
“Generate this scene on four image models in parallel. Show me side-by-side.”
→ The agent fires four parallel POSTs to nano-banana-pro, gpt-image-2, flux-pro/kontext, and ideogram/v3, then returns the grid.
The catalog
Active models
50
Across image, video, audio. Catalog at /models-catalog.json.
Video models
17
Wan 2.7, Veo 3.1, Kling v3 Pro, Seedance 2.0, HeyGen. Up to 1080p, up to 12 seconds.
Image models
26
Nano Banana Pro, FLUX Kontext, FLUX 2, GPT Image 2, Ideogram v3, Bria, Topaz, Reve, plus first-party Runflow Solutions.
Public endpoints
94
Customer OpenAPI surface. Verify at docs.runflow.io/api/openapi.public.json.
The artifact
Plain markdown, plain HTTP. Your agent reads three things: a decision rule, a request shape, and an error map. Then it makes calls.
We publish a sha256 alongside the file so your CI can pin it. When we ship a new version, the hash changes and your agent picks up the update on the next session.
Read the full skill1. Browse https://www.runflow.io/api. If a Solution covers the task, use it. 2. If no Solution fits, fall back to the Model API. 3. Do not reach for the Model API to reimplement something a Solution already does.
POST https://api.runflow.io/v1/models/{owner}/{slug}/runs
Authorization: Bearer $RUNFLOW_API_KEY
Content-Type: application/json
{ "input": { ... }, "callback_url": "https://your.server/webhook" }401: bad/missing key, reprompt user 403: key lacks scope, user adjusts in dashboard 429: rate limited, back off 5xx: transient, check /v1/health, retry
What you get
Plain HTTP, real prices, machine-readable specs. The same shape across image, video, and audio. Every primitive an agent needs to ship a feature, none of the runtime lock-in.
Your agent reads the skill once and makes plain HTTP calls. The same code runs from a script, a serverless function, a CI job, or another agent. The skill works without an MCP runtime.
Real machine-readable public spec at docs.runflow.io/api/openapi.public.json. Generate clients in any language. Pin schemas in your tests. Version drift is detectable.
Calls are routed across providers based on availability and latency. Failover is automatic. The skill itself never tells your agent which datacenter ran the job.
Same auth header, same POST body, same callback contract whether you're calling nano-banana-pro, wan/v2.7, or elevenlabs/tts. The agent learns it once and reuses it everywhere.
Pass callback_url with any run. We deliver a signed payload when it finishes. Signing keys live at /v1/callback-secrets so your handler can verify the call came from us.
Multi-step pipelines you build from primitives: a model run, a check, a fan-out, a callback. Run on a real graph runtime your agent calls directly. No black-box agent flows hidden behind a vendor UI.
BetterPic processed 35 million headshots through Runflow. Their gross margin moved from 40% to 87%. Same product, lower infrastructure cost, less time keeping GPUs alive.
Read the case studyHeadshots shipped
35M+
Gross margin
40 → 87%
Engineers managing GPUs
0
Uptime SLA
99.9%
Common questions
No. The Runflow skill is a markdown integration guide that any agent can read. Once read, the agent makes regular HTTP calls to api.runflow.io. It works with every agent that reads markdown, not just MCP-compatible clients, and you keep portable HTTP code that runs from a script, a serverless function, or a CI job.
Every public customer endpoint in the Runflow API. That's the Solutions API (pre-built workflows like background removal, headshots, outpainting), the Model API (50 active models across image, video, and audio), ComfyUI Deploy (your custom graphs as endpoints), Flows (multi-step composition), and Batch (fan-out with callback). Full reference at https://docs.runflow.io/api/openapi.public.json.
Image: Nano Banana Pro, GPT Image 2, FLUX 2, Ideogram v3, Reve, Bria. Image edit: GPT Image 2 Edit, FLUX Kontext, Nano Banana Pro Edit, Qwen Image Edit, Reve Edit. Video: Wan 2.7, Veo 3.1 (Lite + Fast), Kling v3 Pro, Seedance 2.0, HeyGen. Audio: ElevenLabs TTS v3, Gemini TTS. Plus first-party Runflow Solutions (Background Removal, Outpaint, Object Removal, Skin Fix, Tag Removal, Product Isolation, Model Removal, Background Color). 50 active in the catalog at runflow.io/models-catalog.json.
No package, no plugin, no SDK required. Install SKILL.md into a project or user skill directory, then let the agent read it when Runflow work comes up. Use a short system prompt or rules file only when the agent cannot read skill directories.
Per-call. Each model has a fixed price published in its llms.txt and the catalog. Solutions like Background Removal and Headshots are priced on the Solutions API page. New accounts get $10 in credits to try it.
Multi-cloud GPU pool with automatic failover. 99.9% uptime SLA. BetterPic runs 35 million images a year on this infrastructure through the Runflow API.
Yes. Pass a callback_url with any run. We deliver a signed callback when the run finishes. Signing keys live at /v1/callback-secrets. Your agent does not have to poll.
Yes. Upload a graph at app.runflow.io/deploy. We host it at POST /v1/comfyui-workflows/{owner}/{slug}/runs with the same auth and callback shape as every other endpoint. Your agent invokes it like any other model.
The skill is the opinionated install. llms.txt is the long-form reference. The skill tells the agent how to make decisions (Solutions first, Model API as fallback) and gives a working curl example. llms.txt is a complete catalog of every URL we expose. Most agents only need the skill.
$10 in credits to start. No credit card. The skill works in any agent that can read markdown and make HTTP calls. That's all of them.