Wicket

The control plane for agent context: models, tools, memory and files.

Simulated concept

Style: Swiss grid. Built in HyperFrames from animated SVG, with an instrumental bed and synthesised sound effects. Download the mp4.

Pain point

Companies govern the models and tools that agents use, but nobody governs the memory and files that agents read.

Problem statement

Gateways route model calls. Other tools secure MCP tools. Agents also read memory, files and context. One agent can send a Restricted file to the wrong model, and nobody can show who saw what.

Approach

Wicket is one control plane for models, tools, memory and files. One policy covers all four. An audit trail shows which agent read what, and why. Risky actions go to a person with only the context they need.

Hypothesis

We believe one policy across models, tools, memory and files works better than a separate tool for each. If it holds, a company can show who saw what. To confirm: data is seeded, redaction is a regex, and the audit chain is for show.

Limits
  • All data is a seeded script. No real traffic flows.
  • Redaction is a regex over made-up text, not a real classifier.
  • Memory ages are static. The audit hash chain is for show, not a security control.
Prompts

The prompts that made this prototype, in order. Lightly trimmed.

Step 1 · Claude Code · claude-sonnet-5-5 · 2026-11-12

Using the futureoftheweb.org API key and the video and prototype workflow from earlier prototypes, create 3 to 4 prototypes and launch videos for these themes: Know (Search Beyond the Web), Learning and Attention, Navigating Generated Content, Human-Agent Workflows. Explore an OpenRouter or Runlayer-style product for consumers and enterprise, and lead generation for local service businesses. Each one needs a launch video, an interactive prototype, distinct branding, and a cost report for orchestration, video and prototype. Ask no questions.

Step 2 · Claude Code · claude-sonnet-5-5 · 2026-11-12

Orchestrator brief to each builder: ship one interactive single-file prototype, one 20 to 26 second launch video built in HyperFrames in an assigned style, and a cost report. Hard spend cap of $0.60 per video. No invented statistics. No generated people. Plain words. Look at 12 frames before you report.

Step 3 · Claude Code · claude-sonnet-5-5 · 2026-11-12

Also use Gemini video models on the work Google account for video, and give me 3 to 5 other ways to show each prototype besides a URL and a video.

Step 4 · Claude Code · claude-sonnet-5-5 · 2026-11-12

Wicket looks like separate components, not one unified app with proper UI and UX. I want all of them like that. Rebuilt each prototype as one app: one shell, hash routes, one shared state store, one design system, onboarding, a command palette, and an end-to-end test.

Step 5 · Claude Code · claude-sonnet-5-5 · 2026-11-12

Publish each prototype at futureoftheweb.org/p/ and put the video, the prompts and the tokens on each prototype page as tabs, like the earlier prototypes. Add the other formats too: press kit, spec poster, comic, deck.

Step 6 · Claude Code · claude-sonnet-5-5 · 2026-11-12

The prototypes look squished and not full page. Fix the embed so each one fills the page.

Step 7 · Claude Code · claude-sonnet-5-5 · 2026-11-12

Wicket looks like separate components, not one unified app with proper UI and UX. I want all of them like that.

Step 8 · Claude Code · claude-sonnet-5-5 · 2026-11-12

Update Wicket to target enterprise. OpenRouter is for model inference, but what about memory, files and context? Runlayer is only MCP. I need a 2x2 competitor positioning chart.

Step 9 · Claude Code · claude-sonnet-5-5 · 2026-11-12

Builder brief for Wicket: The control plane for agent context: models, tools, memory and files. Problem: Companies now run many agents. Gateways like OpenRouter route model calls. Tools like Runlayer secure MCP tools. But agents also read memory, files and context, and nobody governs that. One agent can send a Restricted file to the wrong model, and no one can show who saw what. Questions to answer: How might one policy cover models, tools, memory and files together? How might a company see which agent read which memory or file, and why? How might agents hand off to people at the right moment, with the right context and no more? How might one set of permissions follow work across every agent and harness?

Tokens and cost
Cost to make this prototype
StepTotal tokensReal spend (USD)Note
Video and prototype build (sub-agent)185,672$0.30WaveSpeed music $0.30 measured
Extras: press, poster, comic, deck (sub-agent)109,023$0.00
Unified-app rebuild and e2e (sub-agent)237,175$0.00
Enterprise retarget: Memory, Files and Context (sub-agent)289,481$0.00
Total821,351$0.30

Tokens are the sub-agent totals that Claude Code reported. Input and output are not split. The orchestrating session's own tokens were not metered. Token cost is not converted to dollars because the work ran on a subscription.

Real spend is the WaveSpeed music charge. Sound effects, rendering and the prototype itself cost nothing outside the subscription. No Veo or Gemini video call ran.

More ways to see it

Press kit

Share with press or investors

Open

Comic storyboard

Explain the story in a minute

Open

Pitch deck

Present it in a meeting

Open