v1.2.0 opened the door for agents: search, discovery, reading, file upload. But actually running a knowledge base — turning conversation outcomes into notes, maintaining FAQs, organizing the tree, keeping a website in sync — still required falling back to REST. v1.5.0 closes that loop: MCP tools go from 10 to 24, and an agent with a single API key can now run the entire content-operations workflow.
Writing: text and FAQs, no more file disguise
document_ingest_text: write Markdown straight into a knowledge base. Agents previously had to Base64-encode content and disguise it as a file upload — high friction and semantically lossy. Now they write with a title, optionally into a specific folder; with an idempotency key, network retries never produce duplicates.faq_create/faq_update/faq_get: FAQ pairs are a natural memory shape for agents (question → answer). Create, edit and read are all in; editing is guarded by an optimistic revision lock — on conflict, a retryable error code comes back, so concurrent agents never overwrite each other.
Organizing: the file-tree quartet
file_tree_list (nested tree with full paths), file_tree_folder_create, file_tree_node_update (rename/move), file_tree_node_delete (empty folders only). With folder listing in place, the ingest parameter for “put it in this folder” finally works — it had been declared in the schema for a long time, but no tool could obtain a folder ID, so it was silently dropped (also fixed for file imports in this release).
Sourcing: full lifecycle for web content sources
web_source_list / web_source_get / web_source_create / web_source_update / web_source_sync / web_source_delete. Agents used to only be able to append URLs to a single URL-list source; now they can create full-site crawl, sitemap, or URL-list sources, edit crawl configuration, force a full recrawl, and choose whether deleting a source keeps or removes its captured pages. The config shape can be learned directly from existing sources (web_source_get returns the full config), with validation identical to REST.
Zero change to API key semantics
All new tools reuse the existing gating model: what operations a key may perform is decided by its scopes; which knowledge bases it can touch is decided by its binding set. Tool listings and calls are filtered by the key’s scopes — a read-only key sees none of the write tools.