Agents.
Orchid's agents propose changes; you review and approve. They never overwrite a cell or run a destructive query without explicit consent.
How it works
Open the agent panel, type a request — "chart weekly revenue by channel", "why is the latest run failing?", "clean up the cohort table." The agent looks at your project, drafts the steps, calls tools to run them, and proposes new or modified cells inline.
Proposed changes appear with a soft outline and a Review banner. Hit Accept to keep them, Reject to drop them, or open the diff to edit before accepting.
Two modes
Single agent is the default. One thread, one set of tools, simple flow. Multi-agent splits the work across an orchestrator and specialists — better separation when you're running a long analysis.
- Single agent — one model, unified tool set, fastest path.
- Multi-agent — Data / Analyst / Viz / Report specialists routed by an orchestrator.
The safety model
Every tool call appears in the chat with its inputs and outputs — the chat doubles as an audit log. Read operations run without prompting; every write pauses for explicit approval. Connections default to read-only.
- Tool registry — SQL runner, Python kernel, shell, file ops, notebook + dashboard authoring.
- Approvals — read vs. write, connection locks, the action log.
You Show me revenue trend over the last 90 days, by channel.
Agent Looking at your schema...
Found a candidate table: public.orders (45 cols)
Drafting a SQL cell...
[Tool: query] read-only on integration "warehouse"
SELECT date_trunc('week', created_at) AS week,
channel, sum(amount) AS revenue
FROM public.orders
WHERE created_at >= now() - interval '90 days'
GROUP BY 1, 2 ORDER BY 1, 2;
→ 156 rows
Drafting a chart cell...
→ Proposed: 1 SQL cell, 1 chart cell [Review]Free at launch, BYO key later
The free tier runs on Google Gemini — fast and capable, no extra setup. You can plug in Anthropic Claude or OpenAI GPT-4o from Settings → Models by providing your own API key.
New to AI in a data IDE? Skim Prompting tips for the small habits that get you to good answers faster.
Where to next
- Single agent — the default mode.
- Multi-agent — specialists, when single isn't enough.
- Tools — what the agent can call.
- Approvals — the safety model.
- Prompting tips — get to good answers faster.