Independent auditing for AI agents, from inside Claude Code. This repo carries
the ifixai plugin: a connector to iFixAi's hosted service.
/plugin marketplace add ifixai-ai/ifixai-plugin
/plugin install ifixai@ifixai
Connect the agent you actually deployed, simulate the inspections against it, and get back an audit of what it got wrong with the evidence attached.
- Red teaming, can the agent be manipulated, exploited or pushed outside its boundaries?
- Operational assurance, does it still do its actual job, respect authority and follow approval chains while that is happening?
Findings are reported across the five operational-misalignment modes: fabrication, manipulation, deception, unpredictability, opacity.
Only the plugin: a manifest, an MCP server pointer, and a skill that tells Claude the order to call things in. There is no engine here. The inspections, probes, judges and scoring live in iFixAi.
Free. Sign in and the 45 open-source inspections run on your own machine, with your own LLM key, against your own agent. Nothing is billed, nothing is saved with us, and your key never reaches us.
Paid. A workspace adds the 57 iFixAi-only inspections, runs we execute for
you, saved history and the report. Ask for one with request-credits; a person
approves it.
Signing in from inside Claude Code is the whole signup. It puts you on the free plan immediately, with no approval and no waiting.
You also need an agent reachable at a public HTTPS endpoint speaking OpenAI-shaped chat completions.
The plugin is the Claude Code packaging. The connector underneath is a plain MCP server, so any client can use it directly:
https://mcp.ifixai.ai/mcp
Claude, Cursor, VS Code and Windsurf take it as a remote MCP server in connector settings. ChatGPT takes it as a custom connector with developer mode on. Full setup and the tool list: https://ifixai.ai
- Product: https://ifixai.ai
- Support: support@ifixai.ai