This package is the bounded, provider-neutral foundation for the MetaCrate OpenSim grid agent. It contains a reusable library and the metacrate-grid-agent service binary. The default service graph is deliberately offline: it publishes a deterministic ready event, accepts control commands, and shuts down both owned tasks without contacting a grid or LLM. The library also provides the bounded exact-endpoint LLM transport and tool loop for live adapters. The live-grid feature exposes the owner for the existing libremetaverse::GridClient. Its supervised session adapter reuses that client's native NetworkManager for login, event-queue readiness, disconnect notifications, and logout instead of adding a protocol client.

The LLM connection identity has an exact llm.endpoint_url, llm.api_key, and optional llm.model. The endpoint is used exactly as supplied; there are no providers, presets, base-URL rewrites, model catalogs, discovery, or provider SDKs. Debug/Display output removes API keys, grid passwords, URL user information, and URL query values. Secret wrappers are not serializable.

Persistent setup lives in the platform config.yml (normally ~/.config/metacrate/config.yml on Linux). The file contains grid/AI connectivity and the exact privileged avatar UUID list; it is written with mode 0600 on Unix. --preferences opens the setup panel without starting the service. --import-env .env performs a one-time import of the legacy GRID_*/OPENAPI_* values; runtime configuration does not read .env or connection environment variables. JSON files and referenced secret files are accepted only as a legacy migration input.

Run the focused offline gate with:

cargo test --locked -p metacrate-grid-agent
cargo clippy --locked -p metacrate-grid-agent --all-targets -- -D warnings
cargo run --locked -p metacrate-grid-agent -- \
  --config config/grid-agent.example.yml --check-config
cargo run --locked -p metacrate-grid-agent -- \
  --config config/grid-agent.example.yml --run-once

See ../../docs/grid-agent-architecture.md for queue/task ownership, shutdown, and trust boundaries. See ../../docs/grid-agent-llm.md for the LLM wire compatibility envelope, retry rules, and tool-loop safety contract. The central origin matrix, approval binding, budgets, prompt-data boundary, and opaque backend authorization are specified in ../../docs/grid-agent-policy.md. The reconnect state machine, generation fencing, offline-work contract, and shutdown deadline are specified in ../../docs/grid-agent-session.md. Per-avatar/channel expiry, compaction, redaction, optional atomic persistence, and metadata-only operator controls are specified in ../../docs/grid-agent-conversation.md. The versioned integrated/TCP protocol, roles, framing, bounds, TLS remote-mode requirements, events, and management methods are specified in ../../docs/grid-agent-control-plane.md. The unified event schema, pseudonymous correlations, bounded JSONL rotation, fixed-cardinality metrics, diagnostic-capture warning, and non-executing replay contract are specified in ../../docs/grid-agent-observability.md. The embedded/split cross-platform operator interface, keyboard controls, privacy boundary, and focused gates are documented in ../../docs/grid-agent-tui.md. The complete quick start, platform paths, systemd/Windows service operation, secret rotation, backup/upgrade/rollback, failure playbooks, resource defaults, and unsupported-operation list are in ../../docs/grid-agent-operations.md. Milestone acceptance, resource budgets, evidence, and live-grid validation are defined in ../../docs/grid-agent-acceptance.md.