Core problem
Paseo MCP tools only operate on the local daemon. There is no concept of a host parameter at the MCP layer. This means:
- Desktop UI cannot see or interact with remote agents
- Agents cannot manage other agents on remote daemons via MCP
- Subagent tracking does not work cross-host
- The only way to interact with remote daemons is through the CLI
--host flag, which is a workaround
Current state
| Layer |
Local daemon |
Remote daemon |
| CLI |
✅ |
✅ (--host) |
| MCP tools |
✅ |
❌ |
| Desktop UI |
✅ |
❌ |
| Subagent track |
✅ |
❌ |
Proposal
Add an optional host parameter to MCP tools (create_agent, list_agents, send_agent_prompt, get_agent_status, etc.) so they can target a remote daemon (via relay offer URL or direct TCP).
This would unlock:
- Desktop UI showing remote agents — the client already renders agents from MCP; once MCP can reach remote daemons, the UI follows naturally
- Cross-host agent orchestration via MCP — agents can spawn/monitor/communicate with agents on other machines without shelling out to CLI
- Unified subagent tracking — parent agents can track children regardless of which host they run on
- Consistent permission model — remote operations go through the same MCP permission layer as local ones
Use case
Handing off tasks to a GPU/XPU server for compute-heavy work (training, inference, profiling) while maintaining the same interactive experience as local agents — real-time output, follow-up prompts, status tracking, all from the local desktop client.
Environment
- Paseo desktop app on macOS
- Remote daemon connected via relay (prelay.baidu.com:80)
- Relay offer URL:
https://paseoapp.baidu.com/#offer=...
Core problem
Paseo MCP tools only operate on the local daemon. There is no concept of a
hostparameter at the MCP layer. This means:--hostflag, which is a workaroundCurrent state
--host)Proposal
Add an optional
hostparameter to MCP tools (create_agent,list_agents,send_agent_prompt,get_agent_status, etc.) so they can target a remote daemon (via relay offer URL or direct TCP).This would unlock:
Use case
Handing off tasks to a GPU/XPU server for compute-heavy work (training, inference, profiling) while maintaining the same interactive experience as local agents — real-time output, follow-up prompts, status tracking, all from the local desktop client.
Environment
https://paseoapp.baidu.com/#offer=...