Desktop quickstart

The installable Desktop app is the shortest path to creating, observing, and managing research runs. Its distribution contains Runtime, the control surface, the managed plugin installer/worker substrate, and seven reviewed autonomous-research plugins, including the Xsuite and deterministic symbolic-regression capabilities. It also carries the general PySR provider in its local Available catalog. Xsuite and PySR are prepared in isolated plugin environments; other specialized plugins and dependencies are installed on demand. You do not need a repository checkout, a system Python, or a manually started local server.

1. Install and launch

Obtain the package for your platform from the project's release channel: a Windows .msi or setup program, a macOS .dmg, or the native Linux package supplied for the release. Launch Scientific Agent after installation.

The first-run assistant covers:

  1. Welcome — enter setup.
  2. Storage — confirm the Scientific Data Location. It defaults to ScientificAgent under your Documents directory and holds durable runs, models, reports, and artifacts.
  3. Provider — configure and test a model service, or select Continue Offline. Offline mode is sufficient for this deterministic example.
  4. Capabilities — inspect discovered plugins, the local executor, and Codex CLI status.
  5. Ready — enter the new-run page.

If data from an older version is detected, a migration page appears first. You can copy, verify, and move it to the recommended location, keep using the old location, or choose another one. Run IDs, events, provenance, and artifact hashes are not rewritten.

2. Start the first task from one request

On the default New Research Run page, describe the outcome directly. For example:

Analyze the attached data, find candidate relationships, state each formula's valid range and confidence, and deliver a report.

Drop in any datasets, lattices, solver source, or other research files, choose the balanced budget, and start. Runtime profiles the attachments, selects and prepares this project's plugin capabilities, freezes the derived acceptance, budget, and deliverables as a Research Contract, and opens the Control Room. This default path needs a connected Scientific Reasoner; selected plugins from the local catalog are installed and enabled on demand.

If you need a callable model or tool, include its error threshold, domain, fallback behavior, and desired export in the same request. A controlled publisher creates a project plugin only after the plan requires one and the predictor passes its gates; report-only work does not generate a plugin.

3. Optional deterministic offline check

To check Runtime, events, metrics, and the UI without a model connection, open Advanced research contract and enter:

  • Task title: Toy response quickstart
  • Goal: Reach a measured score of at least 0.90
  • Capabilities: select only toy.evaluate
  • Acceptance criterion: score >= 0.90
  • Budget: Actions 10, Experiments 5, Failures 3
  • Scientific workflow: leave the optional switches off
  • Input artifacts: none

Select Review Research Contract. Confirm that the preview names toy.evaluate, score >= 0.90, and the deterministic local Reasoner, then select Start Research.

The reference run normally makes two probes: the first records score=0.64, and the second records score=0.93. CriteriaEngine then sets the run to SUCCESS. This is a deterministic offline product check, not a claim about real scientific-model performance.

4. Inspect the result

In the Research Control Room:

  • filter the timeline to Evidence and expand both experiments;
  • open State and confirm that Criteria progress is satisfied;
  • open Metrics for the two-point score series;
  • open Budget and confirm two actions and two experiments were used;
  • open Lineage, select observation.json, and inspect its SHA-256, creator, content, and relationships.

At the bottom, ask Why did this run succeed?. Ask returns references to recorded events or metrics but does not modify research state.

5. Configure a real model

Open Settings → Providers:

  • If the official Codex CLI is already authenticated, enable Local Codex login. The app does not copy the Codex credential.
  • Remote services may use OpenAI Responses, OpenAI-compatible Chat Completions, or Anthropic Messages formats.
  • API keys go only to the operating-system credential store. Use Test connection after saving.

Then open Settings → Models:

  • Simple Mode assigns one provider/model to scientific reasoning, code generation, and repair.
  • Turn Simple Mode off to assign each role independently.

Settings changes define the default for new runs. For an existing non-terminal run, open Control Room → State → Active models → Switch models. An in-flight call finishes on its old route; the audited update applies from the next model call without discarding research context.

Next steps

results matching ""

    No results matching ""