Build applications around work you can inspect¶
Geyser is useful when your application needs an Agent to do bounded work, return a typed result, and leave a durable record of decisions and effects. Start with one workflow: normalize incoming issues, review supplied material, or evaluate an output before your application accepts it.
This documentation covers SDK/CLI 0.3.0. SDK/CLI 0.1.0 does not contain these execution and reliability corrections. Remote execution requires a matching Customer Cell and Agent rollout; installing a new SDK cannot upgrade a workspace. Check availability before connecting.
Choose a useful first project¶
| Build | What you get | Start here |
|---|---|---|
| Issue intake | Validated issue JSON normalized into a stable application record | Reference applications |
| Output review gate | A deterministic check that every claim cites a supplied source ID | Reference applications |
| Agent document review | A budgeted Open Agent task with a JSON result and inspectable run | Python SDK |
| An existing integration | Idempotent task submission, event cursors, scoped access and result retrieval | Authentication |
| An app on your own model | Calls to a model you taught, running on your Geyser Host, through the OpenAI or Anthropic SDK | Use your own models |
The review gate checks reference coverage. It does not establish that a claim is factually true. Keep the human or domain-specific review your application needs.
Why use Geyser?¶
Use Geyser when you need durable task identity, current workspace authority, inspectable effects and approvals, and customer-controlled data custody together. A direct model API may be enough for a one-shot text transformation. A normal function is simpler for isolated deterministic business logic; packaging that function is useful when you want the same validated bytes tested locally and installed on a customer Agent.
The public SDK and CLI are MIT-licensed. Models, Agent compute, and external tools come from your workspace and may incur charges. Calling your workspace’s own models from your code is free, because they run on your own computer; see Use your own models and costs, limits and support.
First success¶
Run an actual credential-free handler in the quickstart, then create a project in the console’s Developers page. Submit a small task, read its result, and inspect its run before adding your application’s own behavior.
No external adoption, independent certification, or bug-free guarantee is claimed. The repository includes executable examples and regression tests; try the workflow and report where it fails for your use case.