Deploy Foundry
Local Foundry uses fast memory adapters. Production Foundry keeps the same definitions and replaces operational seams: durable data, execution claims, schedules, artifacts, credentials, transports, and telemetry.
Production adapter checklist
- Data: atomic instance provisioning, conversation writes, installation updates, and idempotent inbound claims.
- Execution: bounded concurrency, cancellation, retry classification, leases, and recovery after process loss.
- Schedules: durable once/cron triggers, update and cancellation, plus sleep continuations.
- Credentials: user-owned secret storage, acquisition and refresh with capability-scoped resolution.
- Workspaces: persistent VFS and artifact storage with size, egress, and read-only policies.
- Observability: durable events, redaction, retention, and an exporter appropriate to your environment.
Run without the watcher
pnpm typecheck
pnpm start
# or explicitly
glove foundry start --host 0.0.0.0 --port 4141Package Foundry workers
Build the Foundry package and deploy its start command in your ordinary Node container or process runtime. Mount working environments through Foundry adapters, keep file payloads behind typed artifact references, and give each worker only the capabilities its operational role needs. Glovebox is deprecated and should only remain in existing deployments while they migrate to this model.
Scale by operational role
The application API can remain stateless while adapters own durable data. Split inbound workers, scheduled activators, model execution, and artifact workers when load requires it. Every process uses the same instance reconstruction contract and run correlation, so scale does not introduce a second agent model.
Do not accept traffic before health
The /health response reports discovery, execution, environment, and activation readiness. Treat required adapter failure as startup failure. A deployment that can receive an inbound event but cannot claim or persist it is not healthy.
The cloud path
Foundry’s public contract is intentionally deployment-neutral. A future managed control plane can discover projects, provision adapters, deploy workers, and aggregate inspection without changing definitions or leaking infrastructure primitives into application code. Today, deploy it wherever Node 20 and your selected adapters can run.