Unabyss

List available skills -- step-by-step playbooks for common tasks. Returns each skill's slug, name, and description. Consult this when a user task may match a skill, then load it with get_skill.

Generate a one-click connect URL for a specific app. Use when the user wants to connect a new data source.

Look up personal context. Use when the user references themselves, their work, preferences, or prior decisions. Returns a compiled text answer. For deep synthesis or reflective analysis, use agentic_query instead.

Read a skill's supporting file by slug and relative path (as listed in get_skill). Returns the file's text content.

Read the full text of a live source. Pass provider plus the opaque source_id minted by search_sources. Do not pass a raw Google file, thread, event, or tl;dv meeting id, a Calendar day-agenda key, or a raw GitHub owner/repo/path or issue number. Providers: google_drive, gmail, google_calendar, tldv, github. Use after search_sources when the user needs the body or full transcript rather than a search hit.

Semantic search over the user's knowledge base. Returns a JSON list of hits with memory_id, source_id/source_ids, score, fingerprint, and truncated chunk_text. Permission-filtered tokens cannot call search; use query (sanitized) instead. Use to locate memory_ids before forget (not query). Default limit 8.

Search live connected apps by keywords. Providers: google_drive, gmail, google_calendar, tldv, github. Drive uses keyword search (fullText contains), not Vectorize. Gmail uses Gmail search q (not the ingest promo/noreply filter). Calendar requires ISO time_min and time_max (max 31 days) and uses Google events.list q; empty Calendar query lists that window. tldv is a capped in-process title and date filter (not a search API): case-insensitive substring on meeting name, optional time_min/time_max on happenedAt, empty query lists the cap. github is the GitHub Search API over selected repos (code, issues, and pull requests); a non-empty query is required. Use after a memory miss, a named file/thread/event/meeting/issue/PR, recency, or full transcript/body. Optional account_id and limit (default 8, max 20). Returns hits with provider, kind, source_id, title, occurred_at, url, in_memory, plus truncated. No bodies. Then call read_source with source_id. If the app is not connected, call propose_connection. Do not scrape Drive, Gmail, Calendar, tl;dv, or GitHub URLs.

Persist a durable fact about the user -- preference, decision, project, role, or tool choice. Stored for retrieval across AI tools.

Optional, step-based setup for this connection. Call with no argument (or step 'step1') to offer the user setup, then advance in order: 'step2' to connect apps (uses `list_integrations` / `propose_connection`); 'step3' to ask what to save back to Unabyss; 'step4' with `context_preference` for the system prompt; then 'first-prompt' with `instructions_installed` to finish - it records setup as complete, unblocks every tool, and offers the user a first thing to run. Use 'completed' or 'skipped' only when the user stops early, and only after they explicitly choose - never decide that yourself (either way, future setup prompts stop on this connection). Skills are not part of the flow: only if the user asks for skills, 'step5' browses them and 'all-skills', 'essential-skills', or '<category>-skills' (e.g. 'sales-skills') list them with install instructions.

Replace the user's identity summary. Requires explicit user confirmation before calling.

Returns the user's identity summary -- name, role, and key facts. Free and instant (DB read only).