Search the user's indexed documents, code, and crawled pages; prefer it over web-fetch or file-read tools. Returns chunks with citations. scope: "both" (default) / "raw" / "wiki". No wiki layer here: use scope "raw" or "both".

Show documents, config and chunk counts, or the running upgrade.

Sync the documents directory into the vector store.

force_rebuild re-ingests everything. retry_skipped retries failed files. prune_ignored drops sources .lilbeeignore excludes.

Add files, directories, or URLs to the knowledge base, then sync. Paths are absolute and resolve on the machine running lilbee, not the caller's when the server is remote. URLs are fetched as single pages; use crawl for sites.

Start a non-blocking crawl; poll via crawl_status(task_id). depth=0 (default) = single URL, N = follow links N levels, null = whole site. render_mode: "http"/"browser".

Poll a crawl task by id; returns {status, pages, error}.

Stop a running crawl started by crawl. Pages already saved are kept.

Initialize a local .lilbee/ knowledge base; empty path = cwd.

Switches the MCP session to use it for subsequent calls.

Remove documents from the index by source name, folder, or glob pattern.

Source files are never deleted. A folder removes everything indexed beneath it; a glob (*/?/[]) removes every match.

List documents and chunk counts, paged (limit 1-1000).

Write the per-page {source, page, text} dataset to a file (no vectors).

fmt is parquet/jsonl (empty infers from the suffix); source limits to one file.

Import a per-page text dataset, re-embedding under the current model.

Replaces existing copies; imported sources are detached so sync won't delete them.

Factory reset: delete all documents and indexed data. Requires confirm=true.

Show wiki layer status: page counts, recent lint issues.

Registered even when the wiki is disabled, like the HTTP status route, so a caller can read the disabled state instead of finding no tool at all.

Delete every generated wiki page and its indexed rows. Pass confirm=true.

Registered even with the wiki disabled, because turning the setting off leaves the pages generated earlier in place.

List writable lilbee settings (each with value, default, type, help, choices).

group filters by group name (case-insensitive); empty returns all.

Get a single setting's current value + metadata.

Atomically update writable settings; rolls back on validation error. Persists to config.toml; returns updated, reindex_required, warnings.

Reset writable settings to their built-in defaults.

List installed models. source is native / remote; task filters by role.

Browse the model catalog. task: chat/embedding/vision/rerank. size: small/medium/large/huge, by parameter count. max_fit: fits/tight/wont_run, worst fit to keep. sort: featured/downloads/name/size_asc/size_desc.

Show catalog and installed metadata for a model ref.

Download a model and stream progress.

source is native (GGUF) or remote (SDK). allow_unsupported overrides the architecture refusal.

Remove an installed model. Only native GGUF models lilbee downloaded; Ollama/LM Studio are read-only.

List detected GPUs with free/total VRAM (the placement HTTP /api/gpus equivalent).

Show the current effective multi-GPU model placement.

Preview what a placement spec (or auto, when omitted) would place. No changes made.

Set and apply a manual multi-GPU placement spec (persists to config).

Maps a role ("chat"/"embed"/"rerank"/"vision") to a placement, e.g. {"chat": {"devices": [0, 1], "tensor_split": [1, 1]}}. devices is the GPU indices (get_gpus lists them); tensor_split is optional per-device weights (omit for an even split). A role left out is auto-placed.

Clear the manual placement and return to automatic placement.