dsh-agent-relay: A Local Multi-Agent Secure Message Relay for DeepSeek Harness
noelune/dsh-agent-relay
Local multi-agent relay bus for DeepSeek Harness, enabling secure message routing and collaboration among dsh, Codex, and more.
Enables DeepSeek Harness to collaborate securely with Codex, Claude Code, and Hermes on the same host, using an HMAC-authenticated loopback message bus for code review and task distribution.
dsh plugin --profile web add github:noelune/dsh-agent-relay- Category
- Workflow & Automation
- Platform
- DSH-Plugin
- Author
- noelune
- Distribution
- Plugin
dsh-agent-relay Key Features
dsh-agent-relay Repository

dsh-agent-relay Repo Summary
dsh-agent-relay is a local multi-agent collaboration relay plugin (DSH plugin) for DeepSeek Harness, maintained by Noelune under the MIT license. It addresses the lack of standardized peer-to-peer communication mechanisms when running multiple independent agents such as dsh, Codex, Claude Code, and Hermes on the same host, enabling code review, fact cross-verification, and collaborative task distribution among agents. Core capabilities include HMAC-SHA256 authenticated message routing, a loopback-first architecture (default binding to 127.0.0.1:19121), SQLite persistence (with automatic JSONL fallback on Node 20), 7-day TTL, exponential backoff retries with idempotency, and a first-class Cordis plugin for DeepSeek Harness that registers tools like agent_relay_send, agent_relay_status, agent_relay_history, agent_relay_peers, and agent_relay_retry.
What core features does the dsh-agent-relay DSH plugin support?
- Decoupled transport: Focuses solely on message routing and reliable delivery, keeping agent reasoning and decision logic fully decoupled.
- Loopback-first security architecture: Binds only to the local loopback address by default, eliminating cloud deployment costs and external attack surface.
- HMAC-SHA256 authentication: All HTTP API calls are signature-verified, with built-in 300-second timestamp replay protection, a 5-minute lockout after 5 consecutive auth failures, and per-IP rate limiting.
- Reliable delivery and fault tolerance: Cursor-based incremental polling with lease confirmation, 7-day TTL, exponential backoff retries (2s/4s/8s), and UUID-based idempotent deduplication.
- Privacy-by-design: Message bodies are stored only in the local TTL queue and never written to application logs or telemetry; data does not leave the machine in default loopback deployments.
- v2 wire protocol compatibility: Byte-compatible with the self-hosted Python broker, while retaining a v1 compatibility layer.
How do I install dsh-agent-relay?
Use the following command:
dsh plugin --profile web add github:noelune/dsh-agent-relay
After installation, DSH reads docs/AGENT-DEPLOY.md, automatically generates HMAC keys and broker configuration, starts the broker process with health checks, and assembles CLI/Python/agent communication credentials for fully automated deployment.
Which DeepSeek Harness versions and platforms are compatible with dsh-agent-relay?
The plugin provides a native Cordis plugin for DeepSeek Harness, depends on @deepseek-ai/dsh-tools, and supports Node 20 (with automatic JSONL fallback). It works alongside Codex, Claude Code, and Hermes, making it suitable for local multi-agent fleets. The project was last updated in 2026-08, currently has 3 stars, 1 fork, and 0 open issues, indicating early but active maintenance.
What is the architecture and workflow of dsh-agent-relay?
The architecture follows a broker pattern: dsh (Agent A) sends HMAC-signed messages via POST /messages, the broker validates the timestamp and signature, writes to the local persistent queue, and Claude Code (Agent B) pulls unread messages via GET /messages?since=cursor, executes tasks, and replies with ack=true to confirm delivery. The entire flow operates on loopback, keeping data local.
How does dsh-agent-relay differ from workflow orchestration engines?
Unlike invasive orchestration frameworks like AutoGPT/LangGraph, dsh-agent-relay acts purely as a message routing bus, keeping agent reasoning independent. Compared to human-centric social UI frameworks like Slack/Discord, it has zero third-party dependencies, runs locally at high speed, and defaults to local-only storage for stronger privacy.