Glove Foundry

Foundry is the application framework for Glove agents. You write each agent as a TypeScript file; Foundry turns the folder into a running service with persisted agents, conversations, a run queue, an HTTP and OpenAI-compatible API, schedules, inbound integrations, approvals and an inspector.

What you get out of the box

When to use Foundry

Use Foundry when the agent is the backend of a product:

• more than one user, tenant or team needs their own agent with its own data and capabilities;
• other services, UIs or SDKs need to call agents over HTTP;
• agents must act later — on a schedule, after a delay, or when a message or webhook arrives;
• people need to approve, redirect or audit what agents do.

If you are adding one assistant to an app you already run and you are happy to own storage and the request path, glove-core with glove-react and glove-next is smaller. You can move the same tools into Foundry later.

The mental model

What happens when a message arrives

1. A message reaches an instance and conversation — from the typed client, /api, /v1/chat/completions, a schedule, or an inbound transmission.
2. Foundry persists the run and queues it on its managed execution daemon, which executes it in a fresh child process.
3. The job reloads the instance and conversation from the data adapter and loads the definition from its file.
4. Each definition field is resolved for this run — literals as-is, functions called with the assembly context (instance, conversation, message, history, installations).
5. Built-in Foundry tools, the definition's own tools, and the apps, shared tools and MCP servers this instance installed are mounted, along with memory, layers, subscribers, the working environment and REPL.
6. The normal Glove agent loop runs; every model pass, tool call and custom event is recorded against the run.
7. The result and transcript are stored; the inspector, event stream and any waiting client see the outcome.

Because step 3 happens on every run, editing an instance (installing an app, changing its context) takes effect on the next message — no deploy, no restart.

The smallest useful agent

import { createAdapter } from "glove-core/models/providers";
import { defineAgent } from "glove-foundry";
import { FileStore } from "../../lib/file-store.js";

export default defineAgent({
  description: "A general-purpose assistant",
  model: () => createAdapter({
    provider: "openrouter",
    model: process.env.OPENROUTER_MODEL ?? "openai/gpt-4.1-mini",
    stream: true,
  }),
  // Where this conversation's transcript lives between messages.
  store: ({ conversationId }) => FileStore.open(conversationId),
  systemPrompt: "You are a precise, practical assistant.",
});

Run pnpm dev, open http://127.0.0.1:4141, and choose Start a run. The tutorial takes this to a multi-tenant support agent with installed tools, schedules, approvals and an OpenAI SDK client.

What Foundry owns and what you own

Foundry never performs OAuth, stores secrets, or picks an account for a user. It records only account metadata and an opaque accessRef that your adapter resolves. See the credential boundary.

Read by task