Tools

Tools are the atomic capabilities the Node exposes to the local Qwen model. They are invoked only when the user's intent matches; OCC never calls a web tool on its own initiative, and it never reaches outside the workspace folder for file operations.

For composite multi-step capabilities that orchestrate several tools together, see Skills.

Web tools (gated)

Web tools are stripped from the tool list unless the user explicitly invokes the web — OCC's design rule is that knowledge should come from community-approved packs, not from arbitrary web pages. The web is reachable, but only on demand.

`web_search`

Search the web (DuckDuckGo backend) for current information. Triggered when the user explicitly says "search the web", "look online", "browse the web", "latest news", or similar phrases.

`fetch_url`

Fetch and read a specific web page. Triggered by a URL in the message or phrases like "open this page", "read this URL". Two-stage extraction: high-quality main-content extraction first, with a crude HTML-strip fallback when the page returns nothing usable (paywalls, JS-heavy SPAs).

explicit intent only

Asking "what is the Inca road system?" runs deliberation against the knowledge packs — no web. Asking "search the web for recent news on the Inca road system" enables web_search. The trigger is in the phrasing, not in the topic.

File workspace tools

OCC has a workspace/ folder inside the OCC directory. The local model can read and write files there, sandboxed — file operations cannot escape the directory. Useful for letting OCC produce files on your behalf or operate on files you placed there manually.

`list_files`

List files and directories in the workspace.

`read_file`

Read a file from the workspace.

`write_file`

Write or create a file in the workspace.

`run_code`

Execute a Python snippet with the workspace as working directory. 30-second timeout. Output (stdout + stderr) is returned to the model.

Document and audio readers (upload folder)

Files attached via the paperclip icon land in the upload/ folder. The model reads them with these tools:

`read_pdf`

Extract text from a PDF file. Uses PyMuPDF. For scanned pages with no extractable text, returns a clear "possibly scanned" message rather than empty output.

`read_docx`

Extract text from a Microsoft Word .docx file (paragraphs and tables).

`read_xlsx`

Extract cells from a Microsoft Excel .xlsx file. All sheets, rows returned as tab-separated values.

`transcribe_audio`

Transcribe an audio file (mp3, wav, m4a, ogg, flac) using OpenAI Whisper running locally. Returns the detected language, the duration, and the transcript.

binary upload behavior

When you attach a binary file, OCC automatically stays in chat mode for the rest of the conversation so these atomic tools remain reachable for follow-up questions. The matching tool is called as soon as you reference the file.

Pack catalog tools

These let the user browse and download knowledge packs from chat, without leaving the conversation.

`list_packs`

Browse the broker's catalog tree at a given path. Returns one level of children, with manifest summaries when the children are packs, and a count of packs reachable underneath when they are categories. Empty prefix lists the top level.

Trigger phrases: "what packs do you have", "browse the catalog", "show me what's under marketing".

`download_pack`

Download a pack from the broker into the local expert-packs/ folder.

safe by default

A downloaded pack is saved DISABLED so it does not interfere with current retrieval. The user activates it from Settings → Knowledge Source. Activation triggers the local search index build automatically.

If a pack of the same path already exists locally, the tool first returns a conflict message and waits for the user to confirm before overwriting — a two-step safety to prevent silently wiping a pack the user may have edited with Forge.

Tool indicators in the UI

When a tool runs as part of an answer, a small monospace tag appears next to the routing badge — WEB, URL, PDF, VERIFY, CATALOG, and so on — so you can see at a glance which tools the model actually used.

How tools combine with deliberation

Tools live in chat mode: the classifier routes single-action requests (read a file, fetch a URL, run code) to a direct tool call, with no retrieval and no multi-agent pipeline. Deliberate mode runs the Expert / Critic / Synthesizer pipeline against the curated knowledge packs and does not use web tools — that is what gives OCC its provenance guarantee.

When a request spans multiple tools or several reasoning steps (e.g. "do deep research on X", "fact-check this claim", "summarize all my uploaded files"), the classifier routes it to a Skill instead — a higher level orchestration over the atomic tools described here.