# Aperture

> Give agents the web they deserve.

Most websites weren’t made for AI agents. Aperture opens a doorway, so they can use those sites without starting from scratch every time.

- Website: https://getaperture.vercel.app
- Source: https://github.com/SankrityaT/Aperture
- MCP endpoint: https://calm-steel-d31rh.run.mcp-use.com/mcp
- Docs: https://getaperture.vercel.app/docs (markdown: https://getaperture.vercel.app/docs.md)

## The problem

The web wasn’t built for agents.

Millions of applications still depend on forms, buttons, and interfaces designed for people.

Agents can navigate them, but every action takes time, and it’s often hard to tell whether something actually worked.

Aperture gives agents a better way in.

## Explore once. Execute anytime.

1. **Explore.** An agent opens a website and learns how it works. No API or custom integration required for supported interfaces.
2. **Compile.** Aperture turns discovered actions into reusable, structured tools that agents can call directly.
3. **Execute.** Agents run those tools within their permissions, without repeating the entire exploration.
4. **Verify.** Aperture checks the result, records what changed, and helps agents recover when something goes wrong.

## Tested against browser agents.

Fresh agent sessions on the same job, each graded from the system’s own records.

| Aperture vs Browser agent | Aperture | Browser agent |
| --- | --- | --- |
| Time | 70 s | 158 s |
| Tool calls | 18 | 33 |
| Cost | $0.23 | $0.46 |

- **2.3× faster. The same work order, side by side.** Remove two residents and their devices from a gate-access admin app. Aperture: median 70 s, 18 tool calls, $0.23. A browser agent: 158 s, 33 calls, $0.46. Both 5 of 5 correct. Run with Claude Code, Playwright MCP, GateAdmin32 replica.
- **8/10 chose it. Agents pick it on their own.** Ten fresh sessions had a browser and Aperture connected, and the prompt never named either. 8 used only Aperture, 2 used both, none used only the browser. All 10 correct; none pressed Send without being asked. Run with Claude Code, Playwright MCP, 10 sessions.
- **20/20 on WordPress. Real third-party software.** Browser only against Aperture only, on Sonnet and Haiku. Every run correct on both sides. Deleting a user: 38 s against 47 s, 8 calls against 14, about 60% cheaper. Trashing two posts: even (45 s against 42 s). Run with WordPress, Sonnet · Haiku, Playwright MCP.

Our own test setup, small samples, every run reported, including the ones where it didn’t win.

## Agents get the keys. You keep the receipt.

- **Works with existing software.** No need to rebuild the applications you already use.
- **Reusable by design.** Discover a workflow once. Run it again without starting from scratch.
- **Autonomy with boundaries.** Agents act within delegated permissions and escalate when necessary.
- **Know what happened.** Every consequential action gets a result you can inspect, including failures and incomplete recovery.

## Works with the agent you already use.

Aperture connects through MCP. No new agent or workspace required.

Compatible MCP clients include Claude, Claude Code, ChatGPT, Codex, Cursor, VS Code, GitHub Copilot, Windsurf, Gemini, Gemini CLI, Cline, Goose, Kilo Code, Roo Code, Kiro, Amp, Antigravity, Trae, LM Studio, Open WebUI, Cherry Studio, and any other MCP client.

### Connect an agent

Give this prompt to the agent you already use. It installs the server, signs in through the provided flow, lists the tools, makes one read-only call, and asks before anything that writes.

```text
Connect to the Aperture MCP server and run a safe first check.

1. Add it as an MCP server named "aperture" at https://calm-steel-d31rh.run.mcp-use.com/mcp. Use your client's own MCP setup. In Claude Code: claude mcp add --transport http "aperture" 'https://calm-steel-d31rh.run.mcp-use.com/mcp'
2. If it asks you to sign in, use the sign-in flow it provides. Never ask me to paste a password or key into this chat.
3. List the tools it exposes. For each one, tell me what it does and whether it only reads or can change data.
4. Make one read-only call and show me the result. Change nothing.

Stop and ask me before anything that writes, deletes, or sends.
```

## FAQ

### What is Aperture?

Aperture turns websites into reusable tools that AI agents can discover and execute through MCP.

### Who is Aperture for?

Anyone building AI agents that need to work with websites, especially legacy systems without usable APIs.

### Does the website need an API?

No. Aperture explores the interface and learns how to interact with it through a browser.

### How is this different from browser automation?

Instead of reasoning through the same clicks every time, Aperture compiles workflows into reusable tools, with execution checks and recovery where supported.

### Does Aperture work with any website?

Aperture is designed to support a wide range of interfaces. Compatibility depends on the site’s structure, permissions, and available functionality.

### What happens if a website changes?

Aperture can detect when a compiled workflow no longer matches the interface. Its self-healing engine is being developed to repair and revalidate those workflows.

### Can agents make changes without human approval?

That’s the goal. Agents operate within delegated permissions, escalating only when necessary.

### Can Aperture undo an action?

When the underlying application supports a recovery path. Aperture distinguishes recoverable changes from irreversible ones.

### Which agents can use Aperture?

Any compatible MCP client, including Claude, Codex, and Cursor.

## Open the rest of the web to agents.

Not every application has an API. Not every business will build an MCP server.
Aperture bridges that gap.