1. What Are Tools, Really?
Standard language models are limited by their training cutoff. Tools extend model capabilities beyond raw text generation by providing real-time external access to live web searches, web page fetching, deterministic calculation, and remote API execution.
2. Built-in Tools
Keyless live search utilizing a robust relay cascade (Jina AI, AllOrigins, DuckDuckGo Lite, SearXNG, Hacker News Algolia) with Wikipedia API fallback.
Fetches and strips HTML from external URLs, delivering clean markdown or plain text directly into the agent context window.
Evaluates arithmetic expressions with exact mathematical precision, eliminating arithmetic hallucination.
3. Visual Tool Indicators in Chat
When an agent invokes a tool call, the chat bubble displays a animated indicator ("Searching the web..." or "Processing requested tool..."). Tool call outputs are processed in the background and injected seamlessly into the reasoning pipeline.
4. External MCP Servers (Model Context Protocol)
Connect any remote or local service speaking Anthropic's open Model Context Protocol standard:
Supported MCP Transport Types:
- HTTP / Streamable:
http://localhost:3001/mcp - Server-Sent Events (SSE):
http://localhost:3001/sse - WebSocket:
ws://localhost:3001
Connect servers via Settings → MCP Tools & Servers to instantly expose custom database queries, local file operations, or API triggers to AI agents.
5. System Prompt Tool Injection
When tools are enabled, available tools across all connected servers are merged and injected into the model system prompt. The model outputs structured JSON tool blocks when external data is required.
6. Security & Isolation Controls
Untrusted Content Shield: All external web content is tagged with protective boundaries to prevent prompt injection attacks.
Private IP SSRF Protection: Web fetch proxying explicitly blocks requests to private IP ranges (10.x, 192.168.x, 127.0.0.1) to safeguard local networks.
7. Best Practices for Prompting Tools
Web Search: Include clear time qualifiers (e.g. "Search the web for Gemini 3.7 Flash pricing as of 2026").
Web Fetch: Provide complete URLs including https:// protocol headers.
MCP Integrations: Use clear, descriptive names for your custom MCP tools and servers so agents understand when to invoke them.
8. Tools Troubleshooting
Search Disabled? Toggle Built-In Web Search Cascade back on in Settings → MCP Tools & Servers.
MCP Connection Failed? Ensure your local MCP server has CORS headers enabled for browser cross-origin requests.
9. Adding Your First MCP Server — Step by Step
A concrete walkthrough using a local filesystem MCP server as the example (adapt paths and names to your own server):
- Run the server. For the filesystem reference server:
npx @modelcontextprotocol/server-filesystem /allowed/dir(check the project's own docs for current flags). Confirm it is listening — most implementations speak HTTP/SSE on a localhost port. - Open Settings → MCP Tools & Servers. Add a new server entry: give it a short name (
files), pick the transport (HTTP/Streamable, SSE, or WebSocket), and paste the endpoint URL. - Add headers if required. Auth headers (e.g. a bearer token) go here. The header block is masked in the UI; server secrets never belong in the server's tool descriptions.
- Test before prompting. Use the server's test/refresh action if present (or send a trivial prompt like "list the tools you have available") and confirm the tools appear. A green status means the manifest was read — tool calls can still fail at runtime.
- Scope the allowance. Start with read-only tools enabled; grant write-class tools deliberately, one at a time, and only after you have seen read-path behavior in the trace (Part 7).
- Prompt nudge: agents do not always self-select an obscure MCP tool — mention it: "Use the files server to read /allowed/dir/notes.md" the first few times until the pattern sticks in your muscle memory.
CORS note restated for the skeptical: because the browser calls the MCP endpoint directly, the server must answer cross-origin requests (and usually an OPTIONS preflight). If curl works but the app cannot connect, CORS is almost always the gap — not your server logic.
10. Permission Levels in Practice — What Runs, What Asks, What Never Runs
Tool permission levels are enforced at the final dispatch point in code — no agent loop can route around the confirmation gate. Map them to real actions:
| Level | What happens | Examples |
|---|---|---|
| Auto | Read-only, low-risk tools run without interrupting you | Web search, calculator, read-only MCP tools |
| Ask first | A confirmation card appears above the composer; the tool fires only on Keep/Approve | File writes, commits, external actions, server mutations |
| Blocked | The call is refused outright and the model is told it cannot have that capability | Anything you classify as write-class while unattended, destructive operations on shared systems |
External MCP servers can never impersonate built-in tools by naming — built-ins are distinguished by server identity, not by tool name prefix. A server calling its tool github_list_repos is still an external server subject to the external-server rules, not the built-in GitHub connector.
11. The Web Search Cascade — Freshness, Failure, and Fallbacks
Built-in search is a relay cascade rather than a single API — reader services first, then open mirrors, with a Wikipedia API as the floor. What that means in practice:
No single free endpoint is reliable. Layers mean a bad day at one provider degrades gracefully instead of hard-failing your research mid-sentence.
The cascade can surface current pages, but index lag and reader-fetch failures mean "as of today" claims still deserve a link check before you rely on them. Ask for sources: "cite the URLs you used".
The model may answer from training data with a stale-knowledge caveat, or report that search is unavailable. Neither path is a silent wrong answer — watch the activity trace (Part 7) to see which path was taken.
Narrow, dated, noun-dense queries win: "React 19 use() hook RFC status 2026" beats "tell me about React stuff". Quotes around exact API names and error strings improve hit rates dramatically.
12. Prompting Techniques That Yield Reliable Tool Use
Name the tool for the job. "Calculate 17 * 38.4 with the calculator tool" makes your intent explicit. Whether it runs depends on model tool support and your permissions; verify the tool result in the activity trace.
Constrain the search window. "Search only after 2025" and "prefer official docs" materially improve result quality on fast-moving topics.
Demand the receipts. "Show the fetched URL and quote the line" makes hallucinated tool use obvious at a glance.
One tool per clause. Compound asks ("search X and then post the summary and calendar-book the review") tempt the model to skip a step or invent a tool — split them.
Prime the failure path. "If search fails, say so instead of answering from memory" produces honest negatives instead of confident stale answers.