The 6 Tools
The agent picks tools on its own based on the goal. You don't tell it which tool to use, and most of the time you shouldn't try — the planner is good at matching tools to intent.
Tool catalog
| Tool | What it does | When the agent picks it |
|---|---|---|
| 🔎 web_search | Searches the web via Tavily. Returns up to 5 results with title, snippet, URL. | Goals about current events, public knowledge, or research |
| 🌐 web_fetch | Fetches a single URL's text content. | After web_search returned a promising result and the agent wants to read the full page |
| 📈 quote_lookup | Live market quotes via Stooq. Stocks (AAPL), indices (SPX), ETFs (GLD), crypto (BTCUSD), commodities (GC.F). | Goals mentioning prices, tickers, or market data |
| 📁 project_context | Returns the active project's name, system prompt, allowed tools list, and max-agent-steps setting. | Goals mentioning "this project" or the project's terminology |
| 📄 artifact_read | Lists or reads contents of an artifact you've saved in your vault. | Goals like "summarize the document I uploaded" |
| 🧠 memory_recall | Cosine-similarity search over your memory chunks from past conversations. | Goals like "what did I conclude about X last week?" |
Forcing a specific tool
If you want to nudge the agent toward a particular tool, just mention it in the goal:
- "Use web_search to find recent news about X"
- "Look up the current TSLA price" (triggers quote_lookup)
- "Search my memory for what I said about pricing" (triggers memory_recall)
Most of the time you don't need to. "What is AAPL trading at?" already gets quote_lookup reliably.
What's not in the tool set (yet)
The MVP intentionally does not include:
- External writes — no sending email, no posting to APIs, no modifying files, no GitHub PRs, no calendar events
- Authenticated services — no calendar, Slack, GitHub, Drive
- Infrastructure — no shell commands, no database queries, no system access
- Image/audio understanding — the agent's tools are all text-in / text-out
If you uploaded an image, regular Chat mode (which routes to a multimodal model) is the right place. If you need scheduled actions or external integrations, those aren't built.
Permission levels
Each tool has a permission level on a 5-tier scale:
| Level | What | In MVP? |
|---|---|---|
| L0 | Local read (your data only) | ✅ project_context, artifact_read, memory_recall |
| L1 | Web read (public internet) | ✅ web_search, web_fetch, quote_lookup |
| L2 | Generation (LLM-driven media) | (deferred) |
| L3 | External write (send/post/modify) | (deferred — needs permission gates) |
| L4 | Infrastructure (shell, DB) | (probably never exposed) |
The MVP caps at L1. L3 tools require additional permission gates and audit logging that aren't built yet.