Gambit is an agent harness framework with a CLI + runtime that helps developers build accurate LLM workflows by providing the right context at the right time.
Success! Created gambit at ~/code/gambit
npx @bolt-foundry/gambit repl gambit/hello.deck.md
Run example as multiple chat turns
You can explore other examples in gambit/examples
Feed each tool only the facts it needs. Gambit scopes each workflow step to local context, so you don’t blow up token windows or hallucinate outside the source of truth.
Break assistants into typed workflow steps, each with its own prompt or compute action. Swap models per step and reuse those steps anywhere.
Define schemas once, enforce them at runtime, and ship with CLI and simulator traces baked in. Failures show up as structured logs instead of vague LLM output.
Mix LLM prompts and compute blocks freely. Gambit tracks each step’s label, metadata, and downstream dependencies so you always see how the workflow fits together.
Reference a step anywhere and it keeps its schema + guardrails, so cross-workflow calls stay type-safe.
Use `gambit run`, the simulator UI, or embed the runtime. When you’re ready for production, swap model providers or plug into Bolt Foundry’s managed evals.
Your assistant lives in a single file that enumerates each workflow step (we call it .deck.*). Run it locally with the CLI, then reuse the exact same steps in CI or Bolt Foundry’s hosted platform when you’re ready for production.
.deck.* files describe your workflow steps—run them directly with the CLI.
$ npx @bolt-foundry/gambit init$ npx @bolt-foundry/gambit repl gambit/hello.deck.md$ npx @bolt-foundry/gambit serve gambit/hello.deck.mdLearn how to define each workflow step’s inputs, outputs, metadata, and how they link together.
See how the runtime watches each step’s busy/idle/error hooks to keep workflows safe and observable.
Start from runnable workflows—hello world, multi-step agents, handler patterns, and more.
Get early access to enhanced AI tooling and structured prompt workflows, and be the first to know when new features ship.