How tool calls work
Short answer: Every tool the AI runs during a chat is logged as a row in a “Tool Activity” box: a plain-language description of what it did, with the full arguments and how the call was resolved available on click. Nothing runs invisibly.
What shows up when the AI uses a tool?
Build Perch’s AI works through 100 built-in tools: reading a file, searching the web, adding a card, generating an image, and so on. Every call is logged in the chat as it happens, grouped into a “Tool Activity” box for that round of the conversation. The box header shows a running step count (“3 steps”) and gets a ✓ badge once the turn finishes.
Each row has three parts: a step number, a short description of what ran (for example Look up "quarterly revenue" on the web (DuckDuckGo) or Open and read the file "report.docx") and, on the right, a tag showing how it was resolved.
What do the ⚡ ✓ ✕ ↗ tags mean?
- ⚡ Auto: auto-approved under the chat’s current approval tier; no banner was shown.
- ✓ Approved: a banner appeared and you clicked Allow.
- ✕ Not available: the AI tried a tool outside the chat’s current mode. Nothing ran; the row is the record of the attempt.
- ↗ List panel: a card change approved separately, in the Lists panel, rather than here.
How do I see the full arguments a tool call used?
Rows with more to show than the summary line (a long prompt, several file paths, a batch of cards) get a ▸ marker. Click the row to expand it: the full arguments print underneath (long text truncates at 2,000 characters), and if the model filled in an optional reason for the call, an “AI’s stated reason: …” line appears too, labelled as the model’s own claim rather than as fact. A row whose single short argument the description already shows in full (one file path, one search query) has nothing more to reveal, so it doesn’t expand.
Can I approve or deny from the activity log?
No. The activity log is a record; approving happens on a separate banner before the tool runs (see The approval system). The log is what’s left afterward so you can check what happened without scrolling back through the conversation. Reopening an old chat rebuilds the same log from the saved history, not just from what happened live.
Good to know
- Large tool calls (a big file write, for instance) show a “Preparing tool call” indicator with a running kilobyte count while the arguments stream in, so you can tell the model is still working rather than stuck.
- A turn can run several tool rounds in sequence (by default up to 10, raisable to 100 as an explicit setting) before the AI stops or asks again.
- Every call that runs (auto-approved or approved) gets a row. A call you deny doesn’t run and isn’t added to the activity box, though the model is still told you declined it.