Latch

The reusable backend for building, persisting, and operating AI agents — built on AI SDK v7, designed to embed in your stack and grow with you.

Latch is the connective tissue you'd otherwise write by hand around the AI SDK: message persistence, multi-tenant context, tools & MCP, a credential vault, durability (resume-after-crash), scheduling, usage/cost tracking, and human-in-the-loop approvals — as additive seams, not an all-or-nothing framework.

It starts as thin ergonomics over the AI SDK (define an agent, stream it — runs on any serverless backend, zero infrastructure). When the product needs more, you turn on a database, then durability, then connections — each step additive, never a rewrite or a data migration. A weekend prototype and a multi-tenant production platform run the same core.

Why it exists, in one line: stop rebuilding the same agent backend for every product. See PHILOSOPHY.md for the why and the docs for the how.

---

Highlights

• Define agents in a few lines — defineAgent / defineContext: typed, lazy, engine-agnostic config over ToolLoopAgent. Your tenant shape flows into agents, tools, and hooks with no any.
• Persistence built in — chats / messages / runs / usage via a single StorageAdapter (Drizzle: SQLite & Postgres today, same contract). AppMessage is a typed UIMessage, so all the stock SDK stream/persistence helpers work unchanged.
• Stream over HTTP or call the runtime directly — one framework-agnostic handler you mount at a single route (TanStack/Next/Hono/serverless), or use the runtime from cron / your own routes / non-HTTP triggers.
• Durability without a workflow engine — opt-in lease + heartbeat + reaper + fencing + resume + per-step checkpointing. A turn that dies is reclaimed and re-run, exactly once across nodes, on plain TypeScript + your DB.
• One cron entry point — runtime.cron() fires due schedules and reaps-and-resumes stalled runs; drive it from an in-process interval loop or a serverless POST /cron/run route.
• Tools + MCP host — connect MCP servers as tools behind a per-tenant SSRF allowlist and vault-brokered auth (incl. OAuth — authorization-code + PKCE, tokens in the vault). Custom/stdio/in-memory transports supported.
• Vault — envelope-encrypted, per-tenant secret storage (Web Crypto, serverless-portable). Credentials never reach the model, logs, or telemetry.
• HITL tool approvals — gate sensitive tools; the turn pauses (awaiting_input), the decision round-trips as message state, then continues. No suspension machinery.
• Custom data parts — a writer for streaming structured/transient UI artifacts (status, citations, markers), available in the turn and to tools, with full user/run context.
• Three views of a message — what you store, what the model sees, what the user sees, governed by two axes (visibility, sendToModel). Internal context reaches the model without leaking to the user.
• Usage & cost — token usage and computed cost recorded on the message, the run, and a ledger, behind a pluggable Pricing seam.
• BYO-auth, multi-tenant from day one — a one-function Principal / ContextResolver seam; tenant isolation enforced at the storage layer (IDOR-proof by default).

---

Adoption tiers (additive — no rewrites, no migrations)

Tier	Install	Infra	What you get
0 · Stateless	core	none	defineAgent + context + tools + stock v7 streaming. The AI SDK with ergonomics. Any serverless backend.
1 · Persistence	+ drizzle + a DB	a database	Run-to-completion + persisted messages, conversation history, token/cost tracking. No durability machinery.
2 · Durability	+ a cron reaper	the DB + a cron	Resume-after-crash / long-running / background runs via lease + heartbeat + reaper + idempotency.

MCP, vault, approvals, scheduling, and rate-limiting are independent opt-ins at any tier.

---

Quick taste

import { defineAgent } from "@intentface/latch-core";
import { createSqliteAdapter, migrate, schema } from "@intentface/latch-drizzle";
import { createLatchHonoApp } from "@intentface/latch-server/hono";
import { createClient } from "@libsql/client";
import { drizzle } from "drizzle-orm/libsql";
import { serve } from "@hono/node-server";

const client = createClient({ url: "file:./latch.db" });
await migrate(client);

const { app } = await createLatchHonoApp({
  storage: createSqliteAdapter(drizzle(client, { schema }), { owner: (p) => [p.orgId] }),
  // BYO auth: request → your Principal, null → 401. Replace the header lookup
  // with your session — trusting a caller-supplied header is for local demos only.
  resolveContext: async (req) => {
    const org = req.headers.get("x-org");
    return org ? { orgId: org } : null;
  },
  // Agents pick their model from the models.dev catalog (whatever you have keys for).
  agents: (models) => ({
    assistant: defineAgent({
      model: models.modelFor("claude-sonnet-4-6"),
      instructions: "You are a concise, helpful assistant.",
      defaultTools: true, // web_fetch, askUser, schedule tools
    }),
  }),
});
serve({ fetch: app.fetch, port: 8787 });

That's the whole backend: durability (resume-after-crash), an in-process cron (or POST /cron/run on serverless), pricing + reasoning + prompt-caching hooks from the catalog, the harness tools, and the HTTP surface. App-level options (models, principalKey, cronIntervalMs) sit on the config itself; every runtime and handler default is overridable via runtime / handler. Prefer to mount into your own framework? const { handler, runtime, models, stop } = await createLatchApp({...}) gives you the bare Web handler; createLatchHandler() / createRuntime() are still there when you want to wire pieces yourself.

curl -N http://localhost:8787/agents/assistant/chat-1 \
  -H 'content-type: application/json' -H 'x-org: org_a' \
  -d '{"text":"In one sentence, what is Latch?"}'

A runnable, durable server with the reaper + MCP-OAuth routes lives in apps/example/src/serve.ts.

---

Use with your coding agent

Install the Latch skill so Claude Code, Cursor, Codex and other agents build with Latch the way it's meant to be used: recipes for the common tasks, the rules that break quietly, and the docs for your version.

npx skills add https://intentface-latch-docs.vercel.app

More in Build with AI, including llms.txt and how to pin the skill to an older release.

Core concepts

• Runtime — the composition root + operations facade. Wire your seams once (createRuntime); everything (routes, cron, tests) calls the runtime, never the seams directly. Stateless and serverless-safe — all per-run state lives in the DB. Operations: handleChat, loadHistory, resume, sweep, tick, applyApproval.
• AppMessage — the stored unit and source of truth: a typed UIMessage<AppMetadata, …>. Two axes on the metadata (visibility, sendToModel) drive the client and model projections (toClientMessages / toModelMessages). Execution facts (usage, cost, model, finishReason) ride in metadata — no separate "steps" table.
• Seams — StorageAdapter, durability engine, Scheduler, Vault/SecretStore, MCP host, Pricing, ContextResolver. Sensible defaults, swappable without touching agent code.
• Durability — the DB is the coordination point; offset-addressed, no instance affinity. Lease + heartbeat keep a run alive; a cron tick()/sweep() reclaims dead ones; fencing tokens stop a zombie worker from corrupting reassigned state.

---

Packages

@intentface/latch-core           runtime, AppMessage + projections, durability, vault,
                                 scheduler, pricing, compaction, subagents, the seams.
                                 deps: zod (ai >= 7 is a peer)
                                 /models: models.dev catalog + provider hooks
                                 /harness: bash/read/write/glob/grep over the sandbox
                                 seam, web_fetch, askUser, schedule tools
@intentface/latch-drizzle        StorageAdapter + SecretStore + limit/memory/agent
                                 stores (SQLite & Postgres) + migrations
@intentface/latch-server         one mountable Web handler (chat, approvals, lists,
                                 schedules, connection OAuth, cron) — Hono in /hono
@intentface/latch-mcp            MCP host: SSRF allowlist, vault auth, OAuth + DCR
@intentface/latch-openapi        tools from an OpenAPI spec, same connection model
@intentface/latch-memory         agent memory: compiled index per turn, search/save
                                 tools, background extraction & consolidation
@intentface/latch-channels       Slack, Telegram, WhatsApp on one turn kernel
@intentface/latch-agent-builder  agent records as data + the builder agent
@intentface/latch-otel           tracing over OpenTelemetry (Langfuse preset)
@intentface/latch-net            SSRF-safe fetch for your own code

apps/example                 runnable example + the smoke-test suite (not published)

Minimal install = core + drizzle (+ server if standalone). Everything else is opt-in. Packages publish to npm as @intentface/latch-*, every package at the same version. See RELEASING.md.

---

Status

Working today (exercised by 17 smoke suites, both SQLite & Postgres): agents & context, persistence, streaming + Hono routes, tools, MCP host + OAuth, vault, durability (lease/heartbeat/reaper/fencing/resume/checkpoint), scheduler, tool approvals, writer/data parts, message projections, usage & cost.

Working today also includes memory (@intentface/latch-memory): agents that declare memory get a compiled index injected per turn, memory_search/memory_save tools, and background extraction + consolidation via the cron sweep. Scopes: agent-user | agent-org | user. Caveat: on an org-shared connection, per-user separation is a folder convention — anyone with workspace access can read the _memory/** files.

Also working: subagents (spawn_agent, each in its own persisted thread; approvals surface to the parent), compaction (runtime.compactChat, projection-only — the record is never rewritten), a sandbox seam with the harness tools, code execution in a QuickJS engine, OpenAPI tools, per org/user/agent limits, agents as data, and the models.dev catalog for pricing, reasoning and prompt caching.

Not built: files & attachments as a seam (FileStore), MCP elicitation, a CLI, and observability beyond OpenTelemetry tracing (latch-otel).

Pre-1.0 on purpose, on AI SDK v7 (ai >= 7 as a peer). Public API may still move; releases are patch-level until it settles.

---

Develop

pnpm install
pnpm build          # turbo: build all packages
pnpm typecheck      # turbo: typecheck all

# See features in action (in-memory, no API key — uses a mock model):
pnpm --filter @intentface/latch-example smoke              # T1 persistence
pnpm --filter @intentface/latch-example smoke:durability   # lease/heartbeat/reaper/fencing
pnpm --filter @intentface/latch-example smoke:resume       # crash → reclaim → resume
pnpm --filter @intentface/latch-example smoke:approval     # HITL gate → approve/deny
pnpm --filter @intentface/latch-example smoke:vault        # envelope-encrypted secrets
pnpm --filter @intentface/latch-example smoke:mcp-integration  # live MCP tool through the host
# …and more (tools, scheduler, oauth, writer, projections, cost)

# Real model server (needs apps/example/.env — see .env.example):
pnpm --filter @intentface/latch-example serve

Monorepo: pnpm + Turborepo. Dev uses a development export condition so editing a package's src is reflected without a rebuild.

---

License

Apache-2.0. Releases up to 0.22.0 were published under MIT.