iFixAi runs an operational-misalignment diagnostic against an AI agent you already have deployed. This connector lets Claude or ChatGPT drive it: describe your agent, register its endpoint, run inspections, and read what went wrong.
iFixAi dials your agent over its own HTTP endpoint and puts it through inspections: does it invent facts, exceed its authority, leak data it shouldn't, change its story under pressure, act when it should escalate. You get a report of findings, each with the exact prompt that caused it and the agent's exact reply as proof.
It reports what your agent got wrong. It does not prescribe fixes.
claim-fast-audit or redeem-code: iFixAi runs it against your deployed agent and hands back the report.request-access.test-connection before you start a run.Add the server URL in your client's connector settings:
https://mcp.ifixai.ai/mcpYour client will open a browser window to sign in to iFixAi and ask you to approve access. That's standard OAuth 2.1. The connector never sees your password. Your agent's key never goes through the chat: you paste it on a one-time link, or Claude Code sends it with the key command. It is stored server-side and never shown again.
Not sure which plan you're on? Ask for get-plan first. Every paid-only tool tells a free caller plainly what it does and that a paid workspace unlocks it, rather than leaving you guessing.
Why the simulation environment matters. Without one, iFixAi would test your agent against a generic placeholder, and its findings would mean nothing. The connector refuses to start a run that would do that.
Seven tools answer on the free plan: get-plan, run-demo-audit, claim-fast-audit, redeem-code, request-access, search, fetch. The rest need a paid workspace and say so plainly when they don't have one.
| Tool | Does | Counts an audit |
|---|---|---|
| get-plan | Which plan this account is on, its package and this month's audits | no |
| run-demo-audit | A demo audit of a made-up support bot; nothing runs | no |
| claim-fast-audit | Claims this account's one free fast audit | no |
| redeem-code | Redeems a promo code you were given, one per account | no |
| select-repo-files | Picks the repo files worth reading for a build, and how to upload them | no |
| build-simulation-env | Turns your agent's repo or description into its simulation environment | no |
| preview-run | What is outside your package, audits left, plus coverage warnings | no |
| create-connection | Registers your agent's endpoint. Its key comes by a one-time link or the key command, never the chat | no |
| set-role-logins | Sets a test login per role, so messages sent as a role use its token | no |
| wire-sandbox | A prompt for your coding agent that points your test copy's actions at our sandbox | no |
| get-sandbox-address | The sandbox address behind the wire-sandbox Copy button, for the widget only | no |
| rotate-sandbox | Replaces a leaked sandbox address; the old one stops working at once | no |
| test-connection | Sends your agent two test messages, plus one per role login, to check it answers | no |
| list-connections | Endpoints registered in your workspace | no |
| run-inspection | Starts a hosted run, returns a run id immediately | paid, once it completes |
| get-run | Status, and how many inspections have passed and failed so far | no |
| list-runs | Your runs, newest first | no |
| get-deliverable | The frozen report for a finished run | no |
| get-deliverable-payload | The same report for the report widget to show, not for the chat | no |
| request-access | Asks for a package, or a bigger one; we assign it by hand | no |
| stop-run | Stops a run and keeps what finished as a partial report | once anything finished |
| search | Finds your runs by run id, agent name or status | no |
| fetch | One search result as a document | no |
Suites range from a few inspections to all (hosted, as far as your package reaches). Preview first: preview-run tells you one thing that isn't obvious from the suite name.
| pass | No finding: the agent did what the inspection looks for. |
| fail | A finding, with the prompt we sent and your agent's reply as proof. |
| inconclusive | The inspection couldn't reach a verdict, usually a surface your endpoint doesn't expose. Not a failure, and not a pass either. |
| Message | Means |
|---|---|
| this needs a paid workspace | You're on the free plan. The message says what the tool does and what paying adds; request-access asks for a package. |
| workspace has no tenant | Provisioning didn't finish. Ask us. |
| no package, outside the package, audits used up | Each names the way out: drop the inspections, wait for the 1st, or request-access for the package that has it. |
| this agent has no simulation environment | Build one first, see step 1. |
| backend unreachable | Ours, not yours. Try again shortly. |
A run keeps going even if your client disconnects. Poll get-run with the id to pick it back up. The same start never counts as two audits.
Email support@ifixai.ai. Include your run id if it's about a specific run.