🧠
Durable agent memory
Add the Recorder stage to store useful lessons as episodic, semantic, or procedural memory. Search the native memory store through REST or MCP and carry repository context into later missions. Keep decisions, evidence, and follow-up steps with provenance so later agents can search what was learned.
⚖
Git worktree isolation
Implementation missions get separate worktrees and branches. Broad-scope scheduling reserves a vessel exclusively when requested, while normal independent work can run in parallel. Integration can still have conflicts; your chosen landing mode determines when changes reach the target branch.
⚗
Merge queue
Choose LocalMerge, MergeAndPush, PullRequest, MergeQueue, or None. In queue mode, branches are integrated serially against the current target, with configured checks and recorded results. Enable PR auto-merge only when it fits your workflow.
☍
MCP + REST + WebSocket
An MCP server, REST API with OpenAPI on port 7890, and WebSocket events at /ws let scripts and agents work from the same durable state. The 1.x compatibility policy documents its stable surface and experimental exclusions.
🗃
Your database, your choice
Zero-install embedded SQLite by default, scaling to PostgreSQL, MySQL, and SQL Server — the same schema, migrated across all four.
📊
Built-in observability
Inspect mission history, API request activity, model token charts, and daily token heatmaps. Harbor charts show jobs, occupied slots, launch speed, and link health. OpenTelemetry export and the Docker Grafana stack support deeper operational monitoring.
♻
Self-healing reliability
Stall detection and dock repair can relaunch agents within configured retry limits. Eligible mechanical failures can trigger bounded rescue missions. Failures that require judgment stay visible for a person to resolve.
☵
Backlog, objectives & planning
First-class objectives and a ranked backlog with acceptance criteria and linked evidence. Refine an objective with a captain, plan interactively against the repo, then turn a reply into a dispatch with one click.
📋
Fleet actions & vessel health
Import every repo under a directory in one pass, then grade each vessel: drift, stale branches, outdated or vulnerable packages, tests, and CI. Run one command or one captain mission across all the vessels that need it, with per-vessel results.
🔒
Permissions and tenant boundaries
Tenant-scoped access, secret redaction, and CLI permission policies give you control over who can act and how tools are approved. Operator commands have audit records. Mission tools may bypass approval by default; choose policies appropriate to your environment.
⚓
Local or split with Harbors
Run captain processes on the Admiral host, or use a Harbor where your repositories, git credentials, and agent logins live. The Harbor connects outward to a containerized or remote Admiral. Split mode remains experimental.
🔌
Bring your own inference
Use CLI captains or register a supported hosted/local inference endpoint for an ApiEndpoint captain. OpenAI-compatible, Azure OpenAI, Anthropic, Gemini, Vertex AI, Bedrock, and Ollama adapters support provider-specific configuration. Configure model selection and tool access for each captain.
↻
Controlled upgrades
Experimental source rebuild supports configured native Windows deployments with versioned slots, SQLite backup, and rollback controls. Docker installations update by pulling and recreating their containers. See the deployment guides for setup and limitations.
🧭
Persona-based captain routing
Rank preferred runtime, model, Mux endpoint, or capability-tier classes on each persona. Decide whether a busy first choice waits or falls through. Any role can prefer a local or hosted model; the pipeline stays independent of that choice.
📱
Review from your phone
The iOS/Android client provides saved connection profiles, biometric unlock, activity views, and push approval flows. Deny supported requests from a notification; approving opens the full request. Available as builds, with app-store distribution pending.
✈
Remote operations
Your Admiral dials out to Armada.Proxy over an authenticated tunnel; sign in at the proxy, pick a deployment, and use the same dashboard from anywhere. Missions, logs, diffs, and status history stay queryable.
