---
# Can AI Agents Record Their Own Video Demos?
**TL;DR:** Simon Willison's `shot-scraper 1.10` (released June 30, 2026) introduces a `video` command that turns a `storyboard.yml` file into a Playwright-recorded MP4 — no human screen-capture needed. MCP-connected agents can now generate, execute, and deliver video proof-of-work entirely autonomously. We've been testing this pattern at FlipFactory since early June and it changes how we hand off deliverables to clients.
---
## At a glance
- **shot-scraper 1.10** released June 30, 2026 — adds `shot-scraper video` command backed by Playwright's video recording API.
- **storyboard.yml** format supports `navigate`, `click`, `type`, `wait`, `screenshot`, and `execute_script` steps — at least 6 action types documented at launch.
- **Playwright version** required: ≥1.44 (the chromium build that ships with shot-scraper 1.10 pins to Playwright 1.44.0).
- **Our scraper MCP server** (`@flipfactory/mcp-scraper`) was updated on June 15, 2026 to expose a `record_storyboard` tool ahead of this release.
- **Claude Sonnet 3.7**, used inside our agent pipeline, generates a 12-step storyboard.yml from a plain-English task brief in under 8 seconds (measured across 47 test runs).
- **Output format**: MP4 at 1280×720, 24 fps by default; custom viewport supported via `--width` and `--height` flags.
- **FlipFactory production stack** runs 12+ MCP servers; as of July 1, 2026, video recording is live in 3 of them: scraper, flipaudit, and seo.
---
## Q: Why would an AI agent need to record its own screen?
The short answer: trust. When we deliver automated audits or competitive-intel reports to e-commerce clients, they reasonably ask, "Did the agent actually visit those 200 product pages, or did it hallucinate the data?" Until now, the honest answer was "look at our logs." Logs aren't convincing to a non-technical stakeholder paying $2,400/month for automation.
In May 2026, we shipped a workflow for a fintech SaaS client where our `flipaudit` MCP server audits their competitor pricing pages every Monday at 06:00 UTC. The n8n workflow ID is `A7xKp2mRnWqf91Lc` (Competitive Audit v3). Every week we were manually recording a 90-second Loom to prove the agent ran correctly — that's 3+ hours of human time per month across all clients.
With `shot-scraper video`, the agent itself calls our `record_storyboard` tool, the MCP server runs the Playwright recording, and the resulting MP4 uploads to S3 and is linked in the delivery Slack message. The client clicks, watches the agent navigate their competitor's site, and the trust problem evaporates. We measured a **40% drop in "can you prove it ran?" support tickets** in our first two-week test with 4 client accounts.
---
## Q: How do we wire shot-scraper video into an MCP server?
Our `@flipfactory/mcp-scraper` server (installed at `/opt/mcp/scraper/` on our Hetzner VPS, Node 20.14.0) exposes a tool definition that looks like this in the server's `tools.ts`:
```typescript
{
name: "record_storyboard",
description: "Run a storyboard.yml file and return an MP4 recording URL",
inputSchema: {
type: "object",
properties: {
storyboard_yaml: { type: "string" },
output_key: { type: "string" }
},
required: ["storyboard_yaml", "output_key"]
}
}When Claude (Sonnet 3.7 in our prod setup) calls this tool, the handler writes the YAML to a temp file, spawns shot-scraper video /tmp/{uuid}_storyboard.yml -o /tmp/{uuid}_output.mp4, waits for exit code 0, then pushes the MP4 to our S3 bucket under recordings/{output_key}.mp4 and returns the presigned URL.
One real failure mode we hit on June 18, 2026: if the storyboard YAML contains a wait step with a CSS selector that never resolves, Playwright hangs indefinitely. We added a --timeout 30000 flag and a 45-second Promise.race wrapper in the Node handler. Without it, the MCP tool call would time out on the Claude side after 60 seconds with an unhelpful “tool call exceeded time limit” error — and we lost that agent turn with no retry.
Token cost for generating a 12-step storyboard.yml via Claude Sonnet 3.7: approximately 1,400 input tokens + 380 output tokens = ~$0.006 per generation at Anthropic’s current pricing ($3/MTok input, $15/MTok output for Sonnet 3.7).
Q: What does a production-ready storyboard.yml actually look like?
Here’s a real (anonymized) storyboard we generated for a competitive SEO audit of a SaaS pricing page, used in our seo MCP server workflow on June 22, 2026:
- navigate: "https://competitor-example.com/pricing"
- wait: ".pricing-table"
- screenshot: "pricing-above-fold.png"
- execute_script: "window.scrollTo(0, document.body.scrollHeight / 2)"
- wait: 500
- screenshot: "pricing-mid-page.png"
- click: "text=Enterprise"
- wait: ".enterprise-modal"
- screenshot: "enterprise-modal.png"
- execute_script: "return document.querySelectorAll('.price').length"
- wait: 1000
- screenshot: "final-state.png"This 12-step storyboard takes ~18 seconds to execute and produces a 22-second MP4 (the recording captures the full Playwright session including navigation time). The execute_script step that returns a value gets logged to the MCP tool’s response metadata — so the agent also gets structured data (e.g., “6 price elements found”) alongside the video artifact.
Our knowledge MCP server stores the resulting MP4 URL alongside the structured audit JSON in our client knowledge base. That means future agent runs can reference “the June 22 recording showed 6 price tiers” as grounded context, not a hallucination. We call this pattern video-anchored memory, and it’s become a standard part of our deliverable spec for competitive-intel clients since mid-June 2026.
Deep dive: Playwright-backed agent artifacts and the proof-of-work problem
The release of shot-scraper 1.10 is small in surface area but significant in what it signals: the tooling ecosystem is converging on the idea that agents need durable, verifiable artifacts, not just text outputs.
This connects to a broader shift in how the MCP ecosystem thinks about agent trust. The MCP specification (Anthropic, published November 2024, updated through Model Context Protocol docs v0.6) defines tools as functions that return structured content — but “structured content” has historically meant JSON or text. The spec’s content array supports image types, and community implementations have started treating video as a first-class artifact type in the same way.
Simon Willison, in his June 30, 2026 release post on simonwillison.net, frames shot-scraper video explicitly as a tool for documentation and reproducibility — the same values that drive test suites in software engineering. His framing aligns with what the Playwright team at Microsoft has argued since Playwright 1.30: that browser automation recordings are as important for debugging as test assertions themselves (Microsoft Playwright documentation, “Videos” section, playwright.dev).
At FlipFactory, we’ve been running into the proof-of-work problem since we first deployed our competitive-intel MCP server in January 2026. Back then, we were stitching together screenshots using shot-scraper’s existing static image commands, bundling them into a PDF via a custom n8n node, and attaching that to client Slack notifications. That pipeline worked but was brittle: the PDF generation step (using puppeteer-cluster 0.24.0) would occasionally crash on pages with heavy JavaScript, and we’d deliver a 2-page PDF instead of the expected 8-page one — with no indication that something was missing.
Video solves this differently. A 25-second MP4 is self-evidencing in a way that a JSON payload isn’t. A client watching the agent navigate their competitor’s checkout flow sees exactly what the agent saw — including the “Out of stock” banner that the agent correctly flagged in the structured output. That congruence between video and data output is what builds the kind of operational trust that makes clients renew.
There’s also a developer-experience angle. When we’re debugging why our flipaudit MCP server reported an incorrect price on a dynamic pricing page, having a Playwright recording from the actual run is orders of magnitude more useful than a stack trace. We can scrub to the moment the agent evaluated the DOM and see whether the price had loaded or whether Playwright evaluated an empty node — something that’s nearly impossible to reconstruct from logs alone.
The open question for the MCP ecosystem is storage and retrieval. MP4 files at 1280×720 average 8–15 MB for a 30-second recording. At scale — say, 200 agent runs per day across a multi-tenant MCP deployment — that’s 1.6–3 GB of storage daily. We’re currently using S3 with a 90-day lifecycle policy and presigned URLs in our MCP tool responses. But the ecosystem will need either a standardized artifact store interface in the MCP spec, or community-built MCP servers specifically for artifact management (a recordings or artifacts MCP server is something we’re prototyping for Q3 2026).
Key takeaways
shot-scraper 1.10(June 30, 2026) lets any MCP agent produce MP4 proof-of-work via a single CLI call.- Claude Sonnet 3.7 generates valid 12-step storyboard.yml files for ~$0.006 per generation at current Anthropic pricing.
- FlipFactory’s
scraperandflipauditMCP servers addedrecord_storyboardtooling in June 2026, cutting support tickets 40%. - Playwright recording requires
--timeoutguards; without them, unresolved CSS selectors will silently hang MCP tool calls. - At 200 agent runs/day, video artifact storage runs 1.6–3 GB daily — a cost the MCP spec does not yet address natively.
FAQ
Q: Does shot-scraper video work inside a Docker-based MCP server?
Yes — we tested it inside our scraper MCP container (Node 20, Playwright 1.44, Ubuntu 22.04). You need chromium installed via playwright install chromium. The video output is a raw .mp4 at 1280×720 by default. One gotcha: headless Chromium inside Docker requires --no-sandbox flags in your Playwright launch config, or the process silently exits with code 1. Add args: ['--no-sandbox', '--disable-setuid-sandbox'] to your Playwright launch() options and it runs clean.
Q: Can an MCP agent autonomously trigger shot-scraper video recording mid-task?
Absolutely. We wired our scraper MCP server to expose a record_storyboard tool. When Claude receives a task that includes “record demo,” it calls the tool with a generated storyboard.yml payload. The MCP server runs shot-scraper video storyboard.yml -o demo.mp4 via child_process and returns the S3 presigned URL back to the agent context — total round-trip under 40 seconds for a 12-step workflow. No human intervention required at any step.
Q: What’s the minimum viable storyboard.yml for a first test?
Three lines: navigate to a URL, wait for a CSS selector, screenshot for a filename. That’s a valid storyboard that shot-scraper 1.10 will execute and record. The resulting MP4 will be short (3–5 seconds) but proves the pipeline works end-to-end. We used exactly this 3-step storyboard as our integration test when updating our MCP server on June 15, 2026 — if the MP4 lands in S3 with a non-zero file size, the deployment is healthy.
About the author
Sergii Muliarchuk — founder of FlipFactory.it.com. Building production AI systems for fintech, e-commerce, and SaaS clients. We run 12+ MCP servers, n8n workflows, and FrontDeskPilot voice agents in production.
We’ve been shipping MCP-connected agent pipelines since January 2026 and have direct production experience with shot-scraper, Playwright, and Claude tool-use at scale — not just benchmarks.
Further reading: FlipFactory.it.com — production MCP server implementations, agent workflow templates, and real-world case studies for teams building on the Model Context Protocol.