From a snippet to reproducible evidence

This walkthrough uses a JavaScript scratchpad and needs no network request or project setup. The editor, run, error, recovery, and latest-run Capsule export work on the Free tier in the browser or desktop app. The browser needs a first online load. For the CLI steps, install the separate CLI and provide Node.js 24; start with the CLI guide. Go, Rust, and the local MCP server are desktop-specific, not part of this browser demo. Browser file access uses user-granted File System Access handles where supported; desktop uses its native filesystem bridge and watcher.

1. Run and inspect output

Open Lingua in your browser or the desktop app. Use a JavaScript tab in the default Worker runtime. Replace the editor contents with:

const x = 1 + 2; console.log(x);

Choose Run (or press Cmd/Ctrl+Enter). The console should show 3.

2. Cause an error, then recover

Replace the source with this deliberate failure and run again:

throw new Error('demo failure');

Inspect the error in the console. Replace it with the first snippet and run again; the new result should be 3. This is a recovery test, not a claim that all errors can be fixed automatically. Stop cancels a current run; it does not silently replay the previous source.

3. Save one reviewed run

Open Settings → Account → Run capsules → Save JSON for CLI. The browser downloads lingua-run.capsule.json; desktop opens a local Save dialog. This exports the latest run in Free. Expanded history browsing and opt-in source snapshots are paid features; they are not required here.

Open the JSON and check source.content, result.stdout, and the redacted metadata before sharing. Export includes your source code and may include input/output; it is not an anonymous telemetry event. Importing the file back into Lingua opens a preview without executing it.

4. Validate, then explicitly replay trusted code

Open a terminal in the file’s folder. Validation checks structure but does not execute the source:

lingua capsule validate "lingua-run.capsule.json" --json

Only after inspecting and trusting the source, use the separate command:

lingua capsule replay "lingua-run.capsule.json" --json

Replay executes with your OS permissions. Its comparison reports whether the new status and output match the recorded run; it does not promise identical results for network requests, clock time, dependencies, or different machines. See the Capsule CLI guide for the hash and workspace limits.

Repeat the CLI check without opening the app

The repository includes a fixed RunCapsuleV1 example. From a checkout of the repository, run:

lingua capsule validate docs/examples/deterministic-run.capsule.json --json
lingua capsule replay docs/examples/deterministic-run.capsule.json --json

That example records 3\n; replay’s comparison should report matches: true with Node.js 24. The content hash catches inconsistency, not malice: a modified file can carry a recomputed hash. Never replay an untrusted Capsule.

Agent boundary: Desktop’s local MCP tools are read-only and do not execute code. A shell-capable agent using the separate CLI has a different authority; validation is inert, while replay is an explicit execution decision. Read the agent integration guide before giving either to an agent.