Deployment

Choose the deployment shape, then wire providers.

AtlasClaw supports embedded access and a standalone multi-user AI agent layer. Standalone deployments can run on one node or use the minimal v1.0.0 HA runtime when explicit shared-state, routing, and Channel constraints are met.

Modes

Access and runtime modes.

Embedded deployment

Expose the same Agent through an independent menu entry and an optional Context-aware floating assistant. Both share the enterprise-system Cookie identity; only the floating interface needs the strict page-change bridge.

Standalone deployment

Best when the enterprise needs one SSO-backed multi-user AI agent entry point above multiple systems. A single node supports both long-connection and webhook Channel modes.

High-availability runtime

Runs multiple application nodes against shared MySQL and an initialized shared Workspace, with stable node IDs and sticky routing for each authenticated user.

Configuration basics

The runtime is anchored around providers_root.

{
  "providers_root": "../atlasclaw-providers/providers",
  "service_providers": {
    "jira": {
      "cloud": {
        "base_url": "https://company.atlassian.net",
        "token": "${JIRA_API_TOKEN}"
      }
    },
    "smartcmp": {
      "prod": {
        "base_url": "https://cmp.corp.com/platform-api",
        "cookie": "${CMP_COOKIE}"
      }
    }
  }
}
High availability

Make shared state and node ownership explicit.

v1.0.0 provides a minimal HA runtime whose database, Workspace, node identity, and sticky-routing requirements must be met before startup.

alembic upgrade head

ATLASCLAW_ENABLE_HA=true
ATLASCLAW_HA_NODE_ID=<unique-node-id>
ATLASCLAW_RUN_AGENT_HEARTBEAT=false
Runtime boundaries

Channel ownership never moves implicitly.

  • Use shared MySQL; SQLite is not an HA database. Initialize the shared Workspace and run migrations once before application nodes start.
  • Assign every instance a stable unique node ID and configure the upstream proxy to keep each authenticated user's requests on one node.
  • Set `ATLASCLAW_RUN_AGENT_HEARTBEAT=true` on at most one node when singleton Agent Heartbeat work is enabled.
  • Keep each node's process-owned Token Health, Heartbeat state, and working runtime directory local to that node.
  • HA accepts only registered long-connection Channel modes. Webhook modes are rejected, and Channel ownership does not automatically transfer after permanent node failure.
Operational notes

Preserve governance boundaries.

  • Use `providers_root` to load provider folders from the external providers repository.
  • Keep secrets in environment variables, not committed JSON.
  • Embedded access uses the enterprise-system Cookie identity. Independent menu access only needs an Enterprise System route; the floating interface additionally sends normalized path, nonce, and generation.
  • Keep Context resolution and object actions inside AtlasClaw and Providers rather than sending business DTOs or calling Agent and Tool APIs from Enterprise System code.
  • Use webhook mode for fire-and-forget system-to-system invocation with constrained Skills in a single-node deployment; HA rejects webhook Channel modes.
  • Keep target-platform auth and audit in the provider and the downstream platform.

Deep reference on GitHub