Headless and server-side calculation

Univer separates its content models, calculation, and UI. Headless execution runs the content runtime in Node.js; server-side calculation uses univer-rs for compute-intensive work and file conversion. Applications and agents can combine both around the same snapshots.

A Node.js runtime without UI plugins

Register the required core, product-model, calculation, and Facade plugins without menus, toolbars, or DOM UI plugins. Load Sheets, Docs, Slides, Relational Tables, Canvases, or PDFs according to the selected product stack. Sheets and Docs have Node presets; other products use their own plugin combinations.

Content operations run directly in Node.js. Screenshots and layout checks use a separately configured rendering runtime; the CLI’s Render Runtime can drive a headless browser to produce PNGs and layout findings.

Immediate frontend calculation, versioned server calculation

Both the frontend and server have formula engines. The frontend responds to editing; the server calculates a captured content revision when generating a snapshot or processing a file.

• While editing: the browser calculates affected formulas locally, optionally in a Worker, and displays results immediately. Collaboration records and synchronizes content changes.
• When generating a snapshot: the server’s snapshot policy triggers the save workflow. The server captures the content revision for this operation, calculates with univer-rs, then saves that revision’s content and results together.
• When reopening: the frontend reuses valid results in the snapshot, applies later changes, and continues calculating, reducing repeated full calculation.

Content changes synchronize continuously. Server calculation runs as part of snapshot generation, with results saved for that content revision.

Configure initial calculation to match the saved results:

• WHEN_EMPTY: reuse stored formula values and calculate only missing results.
• NO_CALCULATION: skip initial formula calculation when the application supplies the results.
• FORCED: recalculate all formulas when stored results need refreshing.

These options control initial loading. Subsequent editing continues to calculate in the frontend; the server calculates on demand when generating a versioned snapshot or processing files. See formula configuration, collaboration, and file conversion.

Formula engine performance

See the Univer Rust formula engine benchmarks for 10 workbooks, including Rust input loading, stage timings, input-format boundaries and workbook downloads.

Array formulas and spill results

Within the supported functions and formats, Rust follows Excel’s array-formula and spill semantics. It returns calculated values together with the anchor and spill range, including dynamic arrays and declared ranges for legacy array formulas.

For example, =SEQUENCE(3,2) in an empty A1 produces:

The formula remains in A1; results occupy A1:B3. When an array shrinks, old spill values are cleared. A blocked spill reports #SPILL!. The browser, agent, and export pipeline can reuse the calculated snapshot.

See formula support and Excel array behavior for the corresponding function and range rules.

Connect the runtime to an agent

Keep your existing model, conversation, and tool orchestration. AI SDK tools expose Office content operations to that agent:

1. Inspect structure and read the target content.
2. Apply Facade changes and run the required calculations in the headless runtime.
3. Wait for calculation and commit confirmation, then read back the updated state.
4. Check values and structure; add screenshots or layout checks when visual verification is needed.

Use content operations and visual inspection together. For human review, place edits in a Worktree before comparison and merge. Univer CLI demonstrates how to assemble these capabilities into an agent-facing application.

Integration paths

Back to all features