Define agents with tools, instructions, and memory in code, run them from any function with ctx.agent(), and govern them from the dashboard with approval controls and an audit trail.
npm install @trayai/helix-cli@1.13.0 @trayai/helix-sdk@1.13.0defineAgent and defineToolDeclare an agent in one file under functions/_agents/ with its model, instructions, and tools, and define each tool with a Zod input schema and a handler that gets the caller's ctx. The runtime drives the model loop, runs tools, and stops when the model has an answer or hits the limits you set. Agents load automatically on deploy and under helix dev.
Read the docs →ctx.agent(name).generate() and sessionsCall any agent from any function or schedule. Omit sessionId and the call creates a session; pass one back and the agent continues from the last turn. Sessions come in two scopes: project sessions are shared by every execution, and user sessions belong to the signed-in person who started them. You control the message window and TTL per agent.
Read the docs →Set needsApproval on a tool and the agent stops before running it. The call returns with the pending tool calls, you show them to a person, and a second call carries their approve or deny decision. A predicate form lets you gate only on certain inputs, so a refund over a threshold pauses but a small one does not.
Read the docs →helix agent list and helix agent chat CLI commands
Docs →