---
title: "Documentation for coding agents"
description: "Read the guides, prepare a reviewed workflow, and check its recorded outcome."
canonical: "https://agenttwin.si/docs/agents/"
---

# Documentation for coding agents

Read the guides, prepare a reviewed workflow, and check its recorded outcome.

## Choose your coding tool

Open your tool's MCP setup guide, then connect the AgentTwin documentation endpoint below.



### [Claude Code](https://code.claude.com/docs/en/mcp)

MCP setup guide

### [Cursor](https://cursor.com/docs/mcp)

MCP setup guide

### [VS Code](https://code.visualstudio.com/docs/agent-customization/mcp-servers)

MCP setup guide




## Start with the source guides

Fetch the [Markdown index](/docs/llms.txt), then read the relevant page in full. Each guide has a Markdown link and a Copy page button. The [OpenAPI schema](/openapi.json) describes the REST interface.

| Task | Read first |
| --- | --- |
| Connect a real project | [Project setup](/docs/quickstart.md) |
| Try the optional sample | [Sample test](/docs/example.md) |
| Prepare application code | [Workflow preparation](/docs/preparation.md) |
| Check supported operations | [Provider contracts](/docs/provider-operations.md) |
| Inspect a failed run or verify a fix | [Evidence and retesting](/docs/evidence.md) |
| Choose an execution boundary | [Safety and limits](/docs/safety.md) |

## Connect a documentation client

The public documentation MCP endpoint is `https://agenttwin.si/mcp/docs`, using Streamable HTTP. It exposes `search_docs` and `read_doc` without an account. It reads public guides; it does not execute application code or access customer projects.

See the [MCP connection guide](/docs/mcp/) for client configuration.

## Work with a local repository

Use Node.js 24.11 or newer for this setup. The local MCP server runs as a stdio process in the application's repository:

```sh
npx agenttwin mcp
```

Read the [CLI tool contracts](/docs/cli/#coding-agent-tools) before calling tools. Prepare explicit application files, synthetic fixtures, supported operations and correctness checks. Discovery finds candidates; it does not make every workflow runnable.

## Keep the test meaningful

- Review the workflow and fixtures before executing code. The guided `test_workflow` tool uses Docker; reusable `run_scenario` executes directly on the host. Check the tool contract before choosing an execution boundary.
- Check provider operation coverage before choosing a scenario. A detected dependency is not a complete provider replica.
- Inspect assertions, service events and final state. An untriggered fault or unsupported operation is incomplete evidence.
- For a repair, change application source while keeping the original checks and starting state. Rerun and compare the recorded outcomes.

For a real project, follow [project setup](/docs/quickstart/). The [sample test](/docs/example/) is optional practice with an included app.