GENERAL-PURPOSE AGENT PLATFORM v5.0.0-alpha.5

Give agents a goal.
Keep the evidence.

Turn intent into accountable execution. Route mission steps to your agents, verify their work, and inspect the result.

Your agents bring the tools, models, and domain knowledge. Explore their work in Mission Studio, running on your machine.

Run Mission Studio ↓Watch the walkthrough ↗

THE EXECUTION CONTRACT●

MissionGoal · context · constraints · steps

↓

Coordinator → AgentClaims · selection · bounded execution

↓

Verify → ReportIdentity · attempts · results · evidence

Local agents or authenticated remote Sentinels.
One protocol. Explicit outcomes.

Intent inDescribe the outcome you need.

Independent agentsImplementation stays with the agent.

Evidence outA success claim can be verified.

INTRODUCING MISSION STUDIO

From a plan to a result
you can inspect.

A local workspace for real missions. Start with a working preset, validate your plan, and follow each step to its evidence.

Interface preview — run locally for live execution

starlight MISSION STUDIOLOCAL WORKSPACE

ORDER REPORT

The result, with the receipts.

Completed

Summarize the order data
order-analyst · 3 orders · 5500 cents
Source data checked against the computed total
Write the verified order summary
report-writer · Markdown artifact
Saved contents read back and checked

ARTIFACT CONTENTSMARKDOWN

# Order summary

Orders: 3
Total (cents): 5500

01 / START WITH SOMETHING REAL

Two working presets

Turn order data into a checked report, or fetch service health over HTTP and verify the response.

02 / FOLLOW THE WORK

Progress you can inspect

See live step progress, agent results, and evidence. Request cancellation while a mission is active.

03 / KEEP THE CONTEXT

History that stays with you

Browse saved runs, inspect failures, and download reports. Completed steps retain their evidence.

Your machine. Your agents. Your evidence.

Launch from a checkout with npm run studio, then open the local URL printed in your terminal.

Start a mission ↗

This website is a preview. Start Mission Studio on your machine to run missions; your agents control any external requests. Read the Studio guide ↗

SEE THE CONTRACT IN ACTION

Success is only half the story.

Follow real executions through a verified file, a rejected claim, cancellation, and a remote agent.

Alpha.2 walkthrough of captured CLI and SDK results, with synthetic narration. Later releases add progress storage, mission deadlines, event streaming, and HTTP API integration. Playback is paced for explanation; it is not a live execution timer.

Download video ↓

Inspect the recorded runs

CAPTURED EXECUTIONS · READ ONLY

These are the reports behind the video. Run the examples locally to execute your own agents.

Loading captured results…

Full JSON report

Download all reports and transcript ↗

YOUR FIRST MISSION

Open the workspace.
Make it your own.

Use Node.js 22 or newer. These commands run the current alpha from source; the included presets need no API keys.

Open the local URL printed in your terminal, choose a preset, and run it. Inspect the result and its saved evidence in the same workspace.

Explore Mission Studio ↗

TERMINAL FROM A FRESH CHECKOUT

Get Starlight

git clone https://github.com/starlight-protocol/starlight.git
cd starlight

Install dependencies
```
npm ci
```
Launch Mission Studio
```
npm run studio
```

Prefer the terminal? Run npm run demo for a verified order report.

BUILD ON WHAT IS THERE

A runtime with clear boundaries.

What the platform provides

Sequential missions, agent routing, capacity limits, whole-mission deadlines, cooperative cancellation, optional verification, atomic progress checkpoints, searchable run history, mission validation, and an HTTP JSON agent.

Remote agents use a token-authenticated JSON-RPC/WebSocket contract. “Sentinel” is the protocol name for an agent.

What your agents provide

Tools, model calls, planning, domain logic, and enforcement of business constraints. Starlight does not interpret arbitrary goals on their behalf.

Agents are trusted code. File storage preserves progress across process exits; an unfinished step may already have caused an effect. There is no automatic replay or rollback.

GO DEEPER

Documentation, grounded in the code.

Build an agent ↗Local registration, missions, CLI, and remote integration.Reliable missions ↗Progress storage, deadlines, events, and interrupted runs.Connect an API ↗HTTP JSON agents, service-health workflow, and mission validation.Read the objective ↗First principles, product scope, and what comes next.Protocol 1.0 ↗The interoperable wire contract and its semantics.Review the audit ↗Findings, verification evidence, and known limits.Reproduce the demo ↗Scenario assertions, media generation, and captions.Network security ↗Authentication and deployment responsibilities.