AI Copilot.
A chat panel that knows your canvas. Ask what the clusters look like, why a node failed, or which parameter to change; let it propose the next node as a plan you apply with one click and undo with another; or approve a goal and let it run, check and re-plan. Bring your own key — Oreumics does not resell model access.
Setting it up
- 01
Open the AI Chat tab
It lives in the bottom panel next to Run Log and Methods. Each workflow tab keeps its own conversation, so switching workflows switches context. - 02
Add an API key
Click the gear (API Key Settings) in the chat header and paste a key for OpenAI, Anthropic and / or Google Gemini. Keys are stored locally on this machine and sent only to the provider they belong to. A warning icon next to the model picker means the selected provider has no key yet. - 03
Pick a model
The model dropdown lists the current OpenAI and Anthropic generations —gpt-5.6-terraandclaude-opus-5are the defaults — plus Gemini models. The same picker sets the model used for reports; nodes that call an LLM themselves (Auto Annotation) have their own provider and model parameters.
What the Copilot can see
With every message, the app sends the structure of the current workflow — node types, labels, parameters, run status and error messages — together with a compact summary of each finished dataset (shape, obs columns, embeddings, stored keys). When you ask a question that needs numbers, the model calls a read tool that queries the AnnData on your machine and returns a small result: cluster composition, the top markers of a cluster, a gene’s expression by group, QC statistics, DE results, DMPs, motif hits, which resources are installed, and so on.
What never leaves your machine
Raw matrices, per-cell values and your files stay local. What reaches the provider is node names, parameters, and the small summaries or tool results the model asked for — the same things you would paste into a chat by hand. The Copilot is optional; with no key configured, nothing is sent anywhere.
Asking about results
- “What do these clusters represent?” — it reads the top markers per cluster and interprets them.
- “Why did Clustering fail?” — it reads the node error and the upstream states and diagnoses the usual causes (missing Neighbors graph, missing PCA, a key that does not exist).
- “Is my data normalised already?” — it inspects
adata.X, the layers and the stored flags. - “Which workflow am I in?” — it lists the open project, the focused workflow and the other tabs.
- “Which resources are installed?” — it checks the Resource Center before proposing a node that would stop.
It answers in the language you write in. It is told not to guess numbers it can look up, and to use the node catalogue rather than invent a parameter.
Action plans: Apply, Undo, Guided, Auto
Ask for a change — “add doublet removal after QC”, “set the clustering resolution to 0.8”, “label the left branch treated and make it blue” — and the Copilot answers with a plan card instead of editing the canvas silently. The plan is validated against the real parameter schema before the card appears; an invalid plan is rejected as a whole, never half-applied.
| Control | What it does |
|---|---|
| Apply | Applies the plan to the canvas — creates and connects nodes, updates parameters, sets notes and branch labels — and runs it once if the plan includes a run. No verification, no re-planning. |
| Undo | Reverts the most recently applied plan in one step, including plans that Auto mode or Approve & Run executed. |
| Mode: Guided (default) | Every plan waits for your Apply. The model proposes one step at a time so you review each result. |
| Mode: Auto | Safe plans apply and run without a click, and the model proposes whole pipelines at once. Everything still happens visibly on the canvas, and Undo stays available. Switch with the chip next to the input box, per message. |
Destructive actions always wait for you
A plan that deletes a node or disconnects an edge is never auto-applied, even in Auto mode. It shows what would be removed and waits for an explicit Apply. Looking at a node while a plan card is open no longer invalidates the plan; only real changes to the workflow do.
Approve & Run: the agent loop
Plan cards also carry Approve & Run. One approval, and the agent applies the plan, runs it on the visible canvas, reads the run receipts to verify each step’s actual result against what the plan expected, and — if a step failed or the goal is not met — asks the model to revise and continues. The loop is bounded (a round budget on both sides), Stop cancels it at any point, and Undo reverts what it did. Use it for goals that need iteration: “tune resolution until there are about twenty clusters”. The full behaviour is described on the AI Agent page.
Apply versus Approve & Run in one sentence: Apply executes the proposal as-is, once; Approve & Run executes, verifies and re-plans until the goal holds. Auto mode is only a switch that presses Apply for you.
Reports
- 01
Press Report in the toolbar
The app detects which analyses the workflow performed (QC, clustering, markers, trajectory, spatial, bulk DE, methylation, ATAC, communication, activity …) and collects the finished figures, then posts an outline card into the chat. - 02
Shape the outline
Toggle sections, add a note to any of them, pick the model and the format (HTML or PDF), then Generate. - 03
Say what to emphasise
The Copilot asks whether anything deserves emphasis. If you name an analysis that was never run, it says so and stops instead of inventing it. - 04
Generate
Sections are written one by one from the actual results — tool queries, not memory — with an executive summary at the end. PDF is rendered by the app itself and saved where you choose.
The report narrative is model-written; the Methods paragraph in the bottom panel is generated deterministically from templates and does not need a key.
When it does not work
- A warning icon next to the model picker
- No key for that provider. Open API Key Settings and add one.
- “I have no tool to create nodes”
- Execution is off for this session. The in-app Copilot always has it in installed builds; a connected external assistant may not — see Connect an AI assistant.
- The plan card is disabled
- Another run, a project operation or an agent run is in progress. Wait for it, then Apply.
- The answer quotes numbers that look stale
- Re-run the node; the Copilot reads the stored result of the last completed run.
- Nothing appears after a request from an external assistant
- Plans from Claude Desktop, Codex or Claude Code arrive as Apply cards in the AI Chat tab, with a badge on the tab — not as a pop-up.
Read next