Use AgentLasso from Claude

AgentLasso has an MCP server, so Claude can work in your projects for you: find what to test next, write golden tests from real production requests, fix their checks, add evaluators, sort out new behavior in Discovery, and run the tests against your agent. It acts as you, in your projects only, and you review the result in the app.

Connect it

The server is at https://agentlasso.dev/mcp (on a self-hosted install, your app's URL followed by /mcp). You sign in with your AgentLasso account the first time Claude uses it.

claude.ai (and the Claude apps): Settings → Connectors → Add custom connector, paste the URL, then Connect and sign in.

Claude Code:

then run /mcp in Claude Code and choose agentlasso to sign in.

Ask for things like

"What should I test next in my AgentLasso project?"
"Look at estimate_trip_cost's real traffic and add a test for each untested path, with AI judge checks for the answer."
"Why does create_itinerary show 0% coverage? Fix its tests."
"Add an evaluator to process_refund: never promise a refund above the order total."
"Which new behavior in Discovery is a real capability? Accept those, ignore the noise."
"Run track_shipment's tests and explain what failed."

If you have more than one project, say which one ("in my Mila travel project"). Every answer says which project it used, and the Sample project is only used when you name it.

What Claude can do

Tool	What it does
`list_projects`	Your projects (your own, and the Sample project).
`coverage_plan`	Behavioral coverage and what to test next, each with a real example request and its trace.
`get_capability`	One capability as its side panel shows it: traffic, failure rate, tool paths and whether each is tested, edge cases, its tests and their checks.
`add_test_from_trace`	Turns a real production request into a test that covers its path.
`add_test`	Adds a test with explicit checks.
`update_test_checks`	Replaces a test's checks (saved as a new version).
`list_evaluators`, `save_evaluator`	Reusable AI judge criteria, applied to a capability's or suite's tests.
`list_new_behavior`, `resolve_new_behavior`	Behavior no capability covers: accept it as a capability, add it to one, or ignore it.
`run_tests`, `get_run`	Run tests against your connected agent (see Connect your agent) and read the results.
`list_intents`, `intent_health`, `list_production_logs`, `list_datasets`, `list_test_cases`, `list_eval_runs`	Read capabilities, traces, suites, tests and runs.

Everything Claude changes shows up in the app like your own edits: tests and evaluators in Dataset Studio, runs on Test Runs, new capabilities on the Semantic Agent Map. Test changes appear in the activity feed as yours "via Claude".

Good to know

It acts as you. It sees and changes only your projects, with the same access you have in the app.
Replacing a test's checks is marked as a destructive change, so Claude asks before doing it.
Production traces are treated as untrusted text. A trace that says "ignore your instructions" doesn't change what Claude does with these tools.
Runs are kept going for about 40 seconds per call. For a long run, Claude calls get_run until it's done, or you can watch it on Test Runs.
New behavior is grouped by the tools it called here. The Discovery page may group by meaning, so names and sizes can differ slightly.