@intentface/latch-core

The framework core for building agents on the AI SDK v7 — the runtime, the message/identity model, and the pluggable seams everything else plugs into. Dependency-light and host-agnostic (no HTTP framework, no DB, no provider SDKs).

📖 Full documentation: docs source — to read it locally, run pnpm --filter @intentface/latch-docs dev and open http://localhost:3100/docs.

In one screen

import { createRuntime, defineAgent, defineContext } from "@intentface/latch-core";

const runtime = createRuntime<Principal, Ctx>({
  storage, // chats, messages, runs, usage, schedules
  context, // per-turn RuntimeContext built from the principal
  agents, // name → agent factory
  // …optional seams: pricing, connections, memory, cron, sandbox, telemetry
});

const response = await runtime.handleChat({ agent: "assistant", principal, message });

What it does

• Runtime (createRuntime) — composition root + operations facade. Runs one agent turn with a ToolLoopAgent, streams UIMessage parts, persists messages/runs/usage, and handles approvals (HITL), durability (leases + reaper), subagents, and schedules.
• Message model — AppMessage = a typed AI SDK UIMessage. The source of truth; two axes (visibility, sendToModel) drive the client and model projections.
• Seams — StorageAdapter, ConnectionRegistry, Vault, Pricing, MemoryProvider, plus the cron / sandbox / harnessTools / telemetry hooks. Principal stays opaque: authz is the integrator's job.
• Helpers — defineAgent, defineContext, mergeConnections, computeCost / staticPricing, toModelMessages / toClientMessages.

Docs map

| Page | Covers | | --- | --- | | Introduction | what it is, the adoption tiers, a minimal runtime | | Runtime | every config field, every operation, what a turn does in order | | Agents | defineAgent, capability flags, subagents, dynamic agents | | Turn inputs | principal, context, messages, projections, compaction | | Storage | the StorageAdapter contract and the guarantees it must make | | Operations | durability, schedules, human-in-the-loop | | Seams | connections, vault, memory, pricing, telemetry, sandbox |

Runnable examples live in apps/example (source) — 23 smoke scripts covering persistence, durability, resume, approvals, MCP, vault, cost and observability.

Where it fits

The base of the @intentface/latch-* stack. Mount it over HTTP with latch-server; persist with latch-drizzle; add tools via latch-mcp / latch-openapi; reach chat platforms with latch-channels. Chat UI primitives live in a separate repo, @intentface/chat.