Describe the work. Review the proposal.
The Grid agent drafts source and checks the final proposal within a bounded evaluation scope. You inspect the result, its diagnostics and its limits before choosing to apply it. Tasks and teams require a matching new product build; availability depends on the runtime your instance is using.
MODEL "SaaS pricing" # plans, prices, billing cadence plans = ["Starter", "Team", "Scale"] monthly as currency = [29, 99, 299] annual as currency = monthly * 12 * 0.83 # variable cost per seat cac as currency = [40, 120, 480] payback = MAP(cac, monthly, (c, m) => c / m) END MODEL
Illustrative proposal, not a live measurement · validatedClean: true · evaluatedClean: true
Choose your AI. Use it throughout Grid.
Open AI & extensions to connect OpenAI, Claude, Vercel AI Gateway or your model server. In desktop with the Pro AI runtime, you can install a compatible local model instead. One saved choice serves Chat, new Agent tasks and workbook AI functions. Workbook formulas that name an exact model keep that choice.
The same setup finds signed extensions and lets you choose a registry. These controls require the upcoming app and runtime; prepared model packages become installable only after registry publication. A local server address refers to the computer running Grid.
Bring the context you choose.
Add an image, capture a supported window or camera, or compare two to eight moments from a local recording. An image-capable model uses the same saved AI connection. Review the attachments before sending them.
For a live camera, shared window or selected image input, choose Start observation, set the call limit and stop whenever you need. Selected connector inputs must already supply images. Sessions are temporary and read-only; recordings and a shared history archive are not saved automatically.
Managed tasks can also take image attachments and explicitly share them with their review team. Every selected model needs image support. Task recovery records retain original images until you remove the run.
Video uses sampled images. Audio, direct PDF input and native video-model input are not included. See capture, history and live input →
Four tools for grounded proposals.
Author and research tasks can look up functions, inspect their authorized workbook, compile source and evaluate admitted expressions. Explain tasks return text; external executors receive no Grid tools or authority to apply changes.
Function lookup uses the real catalog. Final-source checks are performed by the host, so the model cannot certify its own proposal by asserting that it passed.
- 01lookup_functionLook up real function signatures and descriptions.
- 02inspect_cellInspect a cell in the task’s authorized workbook.
- 03validate_gridCompile the exact proposed source and return diagnostics.
- 04evaluate_gridEvaluate admitted workbook expressions and report the checked scalar outputs.
Syntax + structure.
{
"ok": true,
"cells": 9,
"warnings": [],
"errors": []
}Behavior, observed.
Check the source. Keep the limits visible.
After drafting, the host independently calls validate_grid and, after successful compilation,evaluate_grid on the exact final source. The evaluator checks up to 100 scalar outputs and refuses unsupported effects. A proposal can return with checks unconfirmed; its warnings explain the limit.
Confirmed proposals carry validatedClean: true and evaluatedClean: true in the response. Application remains disabled until both checks pass, the run has stopped and the editor still matches the captured baseline. These checks do not prove that a proposal meets every business requirement.
Same agent. Four ways to talk to it.
Connect an app, use the CLI, call the API or open Tasks. A team can use different providers while sharing one call budget and deadline. Cancellation stops new dispatch; a remote request already sent may still run or incur a charge.
Chat
$Layer selector → ChatDiscuss the current model and review proposed source changes before applying them. Requires a configured provider or compatible local model.
Library
$createGridAgent({ runtimeRpc, provider: 'gateway' })Connect your application to the task service through an authenticated runtime client.
CLI
$gridctl agent submit plan.jsonSubmit a bounded plan, inspect results, send input, wait for completion or cancel by run identity.
HTTP
$POST /api/agent/generateA compatibility route backed by the same task service. Managed teams use /api/agent/tasks.
Tasks
$Grid Agent → TasksSaved profiles, ordered progress, draft/review teams and proposal review. Progress uses polling, not token streaming.
Bring your own agent.
Grid ships an MCP server, so any MCP-capable agent — Claude, your IDE, your own loop — gets the same grounding: function lookup and search across the full catalog, feature explanations, canonical examples, docs — and grid_validate / grid_evaluate, which compile and run candidate models against the real runtime before your agent shows them to anyone.