How it works
- An LLM with a fixed toolbox of analytics functions.
- It never sees your
org_id— the server pins it from your session before any tool runs, so the model can only query your data. - Results are capped; secrets are never returned.
- Runs on a fast routed model (DeepSeek-V4-Flash by default) through Orbitrage’s own gateway, so its calls show up in your analytics too.
What it can do
These tools act on your own organization only.- Read (anyone)
- Act (org-scoped)
| Tool | Answers |
|---|---|
count_calls | How many calls in a range, filtered by model/provider/workflow/user/tool. |
sum_cost | Total spend, baseline, and savings. |
time_series | A metric over time (calls / cost / errors / tokens) for plotting. |
tool_call_volume | How often function-tools were invoked. |
list_models / list_end_users / list_workflows / list_tools | Leaderboards by traffic or spend. |
detect_anomalies | MAD-based anomaly detection on a daily series. |
node_drift | Per-step drift (tokens, latency, refusal rate) within a workflow. |
list_chatbot_sessions | What’s been asked recently (UI + MCP), which tools fired, any errors. |
Pinned charts & history
- Pin any chart to the Customise page, where it re-renders live by replaying the query. Toggle bar / line / donut / table.
- History is kept per org in your browser. Switching orgs keeps conversations separate; “New chat” clears just that org’s history.
Everywhere, not just the dashboard
The same toolbox is available from:- Slack —
@Orbitrageor/orbitrage, with chart images. - The MCP server — Claude Desktop, VS Code, or Cursor.
list_chatbot_sessions.