Tool health
See how each tool your agents use behaved in real conversations and calls: runs, failures, empty results, platform blocks, top errors, and examples.
GET
Tool health
This report answers “are my agents’ tools still working?”. It reads what actually happened when the agent used each tool with real leads, not how the tool is configured. Use it to find an integration that stopped working, such as an expired credential or a search that never finds anything, before your customers notice.
It returns one row per tool, task, and channel, with the problem tools first.
Before you start
- MCP keys need both
workflows:readandleads:read, because examples include lead data. Full-access REST keys need nothing else. - Test leads, sandbox leads, and Playground conversations are not counted in
tools. Addinclude_test=trueto see them apart, intest_tools. - The report is read-only. Calling it does not run any tool.
- Errors return a stable code in
error, such asinvalid_days. Map the code to your own copy.
Choose what to read
Tests before going live
Withinclude_test=true, the response adds test_tools: the same rows, with the same fields and order, built only from Playground conversations and test and sandbox leads. tools keeps counting real leads only, so tests never mix with real traffic. Without the parameter, or with false, test_tools is not in the response.
Use it to check a task’s tools while you try it in Playground, before real leads arrive. The other parameters apply to both lists. Runs from Test a workflow tool are not counted, because they do not happen in a conversation.
What each number means
How to read the results
Limits
- At most 100 rows in
toolsand 100 intest_tools.truncatedistruewhen more matched in either list. - One request reads the most recent activity first. With
include_test=true, test activity is read in the same request. When the window had more tool activity than one request reads,truncatedistrueandwindow.scanned_fromshows the moment the counts start from. Narrow withworkflow_id,tool_name,channel, or fewerdays. argsandresultare cut at 600 characters. API keys, tokens, passwords, and signatures are masked. Lead data such as names and phone numbers is not removed, so handle samples as lead data.
Authorizations
Query Parameters
Response
Tool rows, sorted by problem rate ((failed + empty) / (calls - blocked)) and then by calls, highest first.
Last modified on October 2, 2026
Tool health