AgentTwin

AgentTwin agent tools

Official public integration resources for AgentTwin agent testing. The product implementation repository is private; this repository contains only public agent instructions, skills and read-only documentation MCP configuration.

AgentTwin is Node.js workflow testing for prepared applications and AI agents. It runs those workflows against synthetic service state, injects supported failures and records operations, state changes and assertions for review. It is a developer integration testing tool, separate from unrelated research projects that use the same name.

Skills

• agenttwin-prepare — prepare a real repository workflow with explicit invocation, synthetic fixtures, supported operations and correctness assertions.
• agenttwin-investigate — investigate recorded failures and verify an application repair against unchanged checks.
• agenttwin-replay — replay a captured run, verify a source-only repair against unchanged checks, and record the restored baseline as a reviewed regression.

Install with a compatible skills client:

npx skills add AgentTwin/agenttwin-agent-tools

Review the skills before installing. This does not install the CLI or run application code. The AgentTwin Node CLI is separate.

The public skill detail pages are agenttwin-prepare and agenttwin-investigate and agenttwin-replay. Directory rankings and security assessments are independent.

Agent configuration

AGENTS.md gives coding agents public integration guidance. Cursor users can opt into .cursor/rules/agenttwin.mdc. The portable Agent Plugins manifest is plugin.json, with the documentation server configured in mcp.json. Use your client's supported plugin installer; client support varies.

The public MCP at https://agenttwin.si/mcp/docs supplies search_docs and read_doc. No credentials are required. It has no customer access or execution. The official MCP Registry entry links to the owned developer page and endpoint.

Maintained references

• Developer portal
• OpenAPI
• API authentication
• World preparation
• Operation contracts
• Synthetic sandbox
• API versioning and deprecation

Local workflow execution can access the host filesystem and network. Review code, fixtures and invocation before running. Static discovery is not a runnable test; no production replica, provider parity or automatic repair is promised.

The canonical maintained source for these public files lives in AgentTwin's website integration bundle. MIT licensed; see LICENSE.