What the agent can do: tools, capabilities and the tool loop

The Senzoukria agent has two families of tools, internet and read-only application tools, shown as a capability grid on the AI page; every tool returns short dated text, and none can place an order, arm anything or write a script.

Senzoukria · Documentation · Updated September 2026


Where to find it

Where
AI Agent page → capability grid "What I can do here"; toggles in the model picker under "What Sen may do"
Defaults
Internet access on, App tools on
Tool rounds
6 per question (MAX_AGENT_TOOL_ROUNDS)
Writes
None: no orders, no arming, no script changes

What it does

The capability grid lists six items and whether each is on or off for the active engine: read your chart and modules, search the web and read pages, read your broker account, read the calendar and the news, read your backtests, know your scripts and the autopilot. Five of them follow the "App tools" switch, one follows the "Internet access" switch. The note under the grid names the engine answering (Claude, Groq or the local model) and points to the model picker for toggling.

Internet tools are web_search (DuckDuckGo, up to 8 results) and web_fetch (readable text of a public page, truncated). Claude receives them as server tools when the web is enabled; Groq and local models receive them as client tools executed by the Rust side, since the webview's CSP does not open the web.

Application tools are all read-only and always offered: read_chart (the live context snapshot), read_account, read_calendar (default 48 hours, max 168), read_news (up to 15 headlines), read_backtests (default 5 runs, max 20), list_strategies, read_autopilot, read_notes (up to 8 matches) and read_note. An unknown tool name returns an explicit error rather than silence.

Settings

Capability switches and tool parameters
SettingDefaultWhat it changes
Internet accessonEnables web_search and web_fetch; Claude uses Anthropic's search, others DuckDuckGo through the app
App toolsonEnables the read-only application tools; never an order, never a script change
read_calendar hours48 (1–168)Look-ahead window of upcoming economic events
read_backtests limit5 (1–20)Number of recent backtest and replay runs, newest first
read_notes results8Matches returned with id, title, tags and excerpt
MAX_AGENT_TOOL_ROUNDS6After this the model is told to stop calling tools and answer with what it has

How the tool loop works

A tool turn spans several API calls. The model replies with blocks, some of which are tool_use; the app executes each and produces one tool_result per call, paired by id, in a single user message; the assistant turn is re-sent exactly as received; the cycle repeats until the model answers without a tool. A call that produced no result is still reported, as an error, so the API never receives a message with a missing result. Server-side tool_use blocks already executed by the provider are not re-executed.

Teaching style and limits

  • The system prompt asks the model to teach in four steps: one sentence that lands, a diagram when the concept is visual, what it looks like on this platform, and what the tool does not tell you. Examples come from MNQ, ES, CL and GC.
  • The model can draw senz-viz blocks of three shapes: chart (4–8 bars), steps (max 8) and compare. Drawn numbers are illustrative, never market data.
  • Conventions are stated as conventions: POC tie takes the lowest price, value area 70 % walked two ticks per side, diagonal imbalance with a minimum quantity, CVD anchored at the CME open 17:00 CT.
  • The prompt states that the platform places no orders through the agent and that positioning is a state, not a forecast. The autopilot tool reports armed or not; it cannot arm, disarm or trade.

This page in other languages

Frequently asked questions

Can the agent place trades or change my scripts?
No. Every application tool is read-only. The autopilot tool only reports state, and the scripting wall and the order-routing lock are untouched by the agent.
Why does the model say it has no internet?
Because the Internet access switch is off in the model picker, or the engine cannot run client tools. The capability grid shows the real state of each capability for the active engine.
What happens after six tool rounds?
The app sends a notice telling the model it has used its rounds and must answer with what it learned, stating plainly what is still failing rather than pretending it is fixed.

Keep reading