hermes - 💡(How to fix) Fix Support remote Hermes agent with local tool execution [1 participants]

Problem

I want to use a remote Hermes Agent instance while keeping tool execution on my local machine.

Example:

• Machine A: local client / workspace machine
• Machine B: remote Hermes Agent with my existing skills, memory, sessions, and model/provider config

Today, if Machine A points model.base_url to Machine B's Hermes OpenAI-compatible API server, tool calls run on Machine B because the API server creates its own AIAgent.

This is surprising because normal LLM providers return tool_calls to the client, and the client executes tools locally.

For example:

pwd

should run on Machine A, where my current workspace lives, not on Machine B.

Desired behavior

I want a split-runtime mode:

Remote brain, local hands.

Remote Hermes should provide:

• skills
• memory / session continuity
• reasoning / model config

Local machine should execute workspace-local tools:

• terminal
• file tools
• code execution
• browser / local dev server interaction
• local MCP servers

Why this matters

Running a separate local Hermes instance loses the remote instance's accumulated skills and memory.

Running everything on the remote Hermes host loses access to the local workspace, local services, local database, and local credentials.

This is especially important for development workflows.

Related issue

Related to #11014, but this issue focuses specifically on:

using a remote Hermes agent while executing selected tool calls on the local client.

Possible solution

Add an explicit split-runtime / local-tool-executor mode.

Example conceptual config:

remote_agent:
  url: "http://remote-hermes:8642"
  mode: split_runtime

tool_execution:
  default: local
  local_toolsets:
    - terminal
    - file
    - code_execution
    - browser
  remote_toolsets:
    - skills
    - memory
    - session_search

Expected flow:

1. Local client connects to remote Hermes.
2. Remote Hermes runs the agent loop and loads remote skills/memory.
3. Remote Hermes emits tool calls.
4. Local client executes allowed local tool calls.
5. Tool results are returned to the remote Hermes agent loop.

Acceptance criteria

• pwd runs on the local machine, not the remote Hermes host.
• Remote Hermes can still use its existing skills and memory.
• Skills created/updated during the session can be written to the remote/shared skill store.
• Tool progress clearly shows where the tool is executed.
• Existing API server behavior remains backward compatible.

Contribution

I am willing to help implement this feature if the maintainers agree with the direction.

Preferred branch name for my implementation work:

feat/remote-agent-local-tools

I can start with a smaller first PR that documents/warns about the current behavior, then follow up with a design PR for the split-runtime / local-tool-executor abstraction if maintainers agree.