v0.13.0 — Toolsets
Toolsets — pick what your AI client sees
Most MCP clients load every tool schema into the model's context at the start of every conversation — AdLoop's full catalog costs roughly 18k tokens per session that way, paid before you type a word. 0.13.0 lets you expose only what you actually use:
- Every tool now belongs to one of seven toolsets:
ads,ga4,tracking,gtm,gsc,web,merchant— plushealth_checkandconfirm_and_apply, which are always included. - Set
ADLOOP_TOOLSETS=ads,ga4in your MCP client'senvblock and the server only advertises those sets. Measured:ads,ga4≈ 13k tokens per session, aga4-only client ≈ 2k — nearly 90% less than the full catalog. - Toolsets are per client, not per install: one AdLoop config can serve a trimmed Cursor and a full-catalog Claude Code side by side. Unknown names fail at startup with the valid list, and a regression test forces every future tool to be tagged.
On AdLoop Cloud, the same feature ships per API key — pick toolsets when creating a key in the dashboard and that key's tools/list is trimmed server-side.
A friendlier adloop init
- The wizard now offers the toolset selection and writes it straight into the printed Cursor / Claude Code snippets.
- New optional steps pin a GTM container and a Search Console property — auto-discovered from your Google account where possible, with guided manual fallback — and store a PageSpeed API key. Every step skips with Enter.
- The generated config now always includes the
gsc:/gtm:/pagespeed:sections, so filling them in later needs no guessing of key names.
Sturdier OAuth when scopes grow
- Stored tokens that predate newly added scopes are discarded up front instead of failing at refresh with a cryptic
invalid_scope. - If Google still refuses a refresh, the error explains the actual fix (re-run any tool, approve the updated consent screen).
- Partial consents — unchecking scopes on Google's granular consent screen — now fail once, loudly, naming the missing scopes, instead of storing a token that would reopen consent on every call.
Docs
- Tool counts are no longer hardcoded anywhere in docs or CLI output; the README explains toolsets and their measured context savings.