Jms Agent Bridge — Development skill for Claude Code

jms-agent-bridge

**An MCP bridge that lets Claude Code, Codex and DeepSeek Harness (dsh) drive a heterogeneous GPU/NPU cluster through a [JumpServer](https://github.com/jumpserver/jumpserver) bastion — without breaking the audit chain.**

Point it at your bastion, describe your hosts once, and a single agent session can probe the cluster, push experiment code, launch a Torch-FL run across Ascend and CUDA nodes in parallel, watch it, and pull results back. Every byte still goes through JumpServer's audited SSH gateway.

Claude Code / Codex / dsh ──stdio MCP──▶ jms-agent-bridge (local)
                                          ├─ control plane: JumpServer REST API  (discovery · audit · tickets)
                                          └─ data plane:    ssh/rsync via KoKo   (commands · sync · jobs)

The bridge runs on **your** machine and is launched by your agent: nothing is installed on the bastion, and your credentials stay in your own key agent and config file.

Compact resident mode

An optional [five-tool resident service](docs/compact-service.md) provides `run`, `read_result`, `discover`, and `call`, complete output files, native SSH connection reuse with explicit persistent shells, remote Docker, and hourly REST asset discovery. Install `pip install "jms-agent-bridge[compact]"` and run `jms-bridge serve --compact --transport streamable-http`. The existing tool surface and YAML workflow remain the default.

Requirements

• Python 3.10+
• ssh and rsync on PATH — the bridge shells out to them and inherits your key agent.
• sshpass — only if you authenticate to the gateway with a password rather than a key. Most distros package it; jms-bridge doctor reports whether it is missing.

Install

git clone https://github.com/lvyufeng/jms-agent-bridge.git
cd jms-agent-bridge
python3 -m venv .venv
. .venv/bin/activate
pip install -e .

`jms-bridge` is now the console script at `/.venv/bin/jms-bridge`. Note tha