# Agentic Stress Tester
> Tests an AI agent network on a website against a checklist of rules, using real models with the
> tester's own key. Rules decide pass or fail; no model grades another model.
Live site: https://agentic-stress-tester.pages.dev
## Zero-token checks (no model call, no key)
These run on a network bundle (network.json and catalog.json, optionally manifest.json, mocks.json,
scenarios.json). Run them from a clone of the repository:
- lint-bundle
: pre-flight, errors and warnings
- call-size : characters each agent's model call sends, against a 16,000-character budget
- validate-call --tool [--agent ] [--mode top|partner] [--args JSON]: would this call be refused, and why
- list-checklists [--industry ]: the checklists and what they test
- estimate --checklist [--scope smoke|all|model_compat]: how many model calls a run would take; spends nothing
- check-trace --checklist --dir : verdicts on one saved run
node cli/tools.mjs lint-bundle path/to/bundle
Exit codes: 0 passed, 1 a check found a problem, 2 bad input. Output is JSON on stdout.
## MCP server (local, stdio)
Offers the same zero-token checks to coding agents. Keys never leave the user's machine.
{
"mcpServers": {
"agentic-stress-test": {
"command": "node",
"args": ["path/to/agentic-stress-test/mcp/server.mjs"]
}
}
}
Tools: lint_bundle, export_catalog, call_size, validate_call, list_checklists, estimate, check_trace.
## Runs that spend model calls
A run calls real models with the tester's own key, so it is not offered over MCP. Use the page, or
the command line with --yes:
TESTER_ROUTER= TESTER_API_KEY= \
node cli/run.mjs --target --cases site/checklists/ --models --yes
## Checklists
Industry checklists (hospitality, restaurants, healthcare booking, banking and payments, travel and
transport) live in site/checklists/index.json. Each case names the fault points it tests, with sources.