Fragmented Agent State
Threads, run logs, and tool outputs scatter across apps. The coordination context breaks on every handoff.
AI Command Harbor
Multi-Agent Coordination, Under One Command Harbor.
AgentDock is a local-first control plane for coordinating agent workflows: route providers, supervise execution, preserve thread history, and resume sessions without context loss.
Current providers
Claude CodeCodexOpenCode
Problem / Why AgentDock
Agent workflows fail when coordination, context, and provider state live in separate tools.
AgentDock closes the gap with a dedicated coordination control plane, not another chat wrapper.
Manual Handoff Overhead
Without a controller, switching provider means manual copy-paste and rebuilding execution intent.
Chat UI Is Not Control Plane
Serious multi-agent work needs routing, supervision, and resumable runtime control, not chat-only screens.
Core Capabilities
Built to coordinate agents end-to-end, not just display messages.
Agent Coordinator Core
Central control loop for routing, supervision, and failover across claude_code, codex, and opencode.
Unified Thread + Run Graph
One timeline that merges prompts, tool calls, outputs, and execution state across providers.
Session Resume Engine
Resume active runs with the same orchestrator state instead of replaying the full context window.
Policy-Based Provider Switching
Route by latency, capability, or cost policy while preserving execution metadata and continuity.
How It Works
Three steps from command intake to coordinated execution.
Start the session, let the coordinator route and supervise provider runs, then resume from the same state.
Dock Your Session
Open agentdock.mrpanda.run and start a workspace-bound control session.
Coordinate and Route
The coordinator dispatches tasks to Claude Code, Codex, or OpenCode from one shared run graph.
Resume Anywhere
Reconnect and continue from the same orchestration state across desktop and remote relay sessions.
Architecture
Coordination Core + Secure RelayLocal Agent Coordinator
Task router + run supervisor + thread graph + resume state
claude_code adapter codex adapter opencode adapter
Local Runtime
Execution stays on-device by default
Encrypted relay ↓ Encrypted relay →
Secure Relay Path
Scoped tunnel for remote monitoring and control
Remote Console (Mobile in Progress)
Live session mirror with controlled actions
Agent coordination logic stays local-first. Remote access is opt-in and tunneled through a secure relay path, so orchestration state remains consistent across reconnects.
Current Progress
Coordinator runtime is operational today, while remote mobile control is under active development.
Desktop Core: Working
- • Coordinator loop and task routing are stable
- • Session resume loop is active
- • Cross-provider thread and run continuity are integrated
Mobile Remote: In Progress
- • Remote control surface is being refined for touch workflows
- • Secure relay reliability is under active tuning
- • Beta hardening is focused on reconnect and coordination state sync
Next Roadmap
The next phase focuses on remote collaboration depth, policy routing, and operational telemetry.
Remote Session Sharing
Share scoped live coordination sessions with collaborators without exposing the full local runtime.
Multi-Agent Routing Policies
Define dispatch rules by task class, cost ceiling, latency target, and provider capability profile.
Operational Insights Layer
Runtime analytics for token usage, failover events, retry patterns, and coordination bottlenecks.