Back to directory

dsh-harness-mcp-server

In a curated listMaintenance: Maintained

chushixixin/dsh-harness-mcp-server

Expose DeepSeek Harness agent capabilities as an MCP server (brain=Hermes, arms=Harness)

View on GitHub
$ npm install @chushixixin/dsh-harness-mcp-server

16

stars

7

forks

TypeScript

Language

MIT

License

2026-08-13

Created

2026-08-15

Last push

MIT TypeScript harness plugin that publishes DeepSeek Harness agents as an MCP server: five tools, structured results, per-cwd session reuse — bound to localhost, unauthenticated by default.

DSH integration

Native runtime

Author-claimed

Safety audit

Unaudited

Last verified

2026-11-16

License

MIT

01What can it help you accomplish?

  • Hand one real coding task from an external MCP client to a Harness agent and get a structured result back in the same call

    A result carrying `sessionId`, `assistantText`, `toolCalls`, `toolResults`, `changes`, `verification` and `leftovers`, so the client can judge the run and remember it

    Teams that keep the planning "brain" in a client like Hermes and want Harness's agent and full toolset as the "arms" doing the actual edits

  • Queue a task instead of blocking the client, then collect its result later

    A queued task carrying the task text, memory context and working directory — and the same structured result, polled back through `task_result`

    Clients that have to stay responsive while Harness runs a long refactor or several unrelated tasks in parallel

  • Close the loop between the client's own long-lived memory and Harness's coding across runs

    Memory fed into each task as `context`, and the run's `changes` / `verification` / `leftovers` written back into the client's memory for the next task

    Agents with persistent project memory that need every Harness run to start informed and end remembered

  • Offload a huge refactor or several unrelated tasks to Harness so they don't blow the primary agent's context

    Work done inside Harness's own agent session, with the session reused per cwd — the README estimates roughly 15–20× cheaper than one-shot `dsh headless`

    People who hit context limits in their main agent, or who want parallel execution of unrelated tasks

02How to install into DeepSeek Harness

Prerequisites

  • A DeepSeek Harness workspace — the npm package is loaded as a Harness plugin through a `cordis.yml` patch, and the README says to reference it from that workspace
  • `DEEPSEEK_API_KEY` exported in the shell that starts Harness (`export DEEPSEEK_API_KEY=...`)
  • `bubblewrap` installed on the host when agent bash commands must write files — the `workspace-write` sandbox refuses write commands without it
  • For the from-source option: the repo cloned under `packages/mcp/harness-mcp-server/` inside the Harness workspace (the pnpm workspace matches `packages/*/*`), plus the `tsconfig.host.json` / `tsconfig.base.json` registrations

Installation steps

  1. 01

    `npm install @chushixixin/dsh-harness-mcp-server` — the README's recommended npm option

    $ npm install @chushixixin/dsh-harness-mcp-server

  2. 02

    Reference the plugin from the Harness workspace through the `cordis.yml` patch: it inserts `id: harness-mcp-server`, `name: '@chushixixin/dsh-harness-mcp-server'` and config `http: true`, `port: 8090`, `host: 127.0.0.1`

  3. 03

    Start Harness with that patch: `corepack pnpm dsh web --patch ./packages/mcp/harness-mcp-server/cordis.yml`

  4. 04

    From source instead: clone into `packages/mcp/harness-mcp-server/`, run `corepack pnpm install`, register the package in `tsconfig.host.json` references and `tsconfig.base.json` paths, then `corepack pnpm exec tsc -b packages/mcp/harness-mcp-server` and `corepack pnpm run build:lib:host`

  5. 05

    Point the MCP client at `http://127.0.0.1:8090/mcp`; for Hermes the README's command is `printf 'n\nY\n' | hermes mcp add harness_plugin --url http://127.0.0.1:8090/mcp`

Verify the integration

  • The README's stated sign of life is the MCP server listening on `127.0.0.1:8090` (StreamableHTTP)
  • Call the `echo` tool — the README lists it as the way to verify MCP connectivity
  • `harness_list_tools` should return Harness's registered tool names

03DSH integration and capability boundaries

DSH integrationNative runtime

DSH (Cordis) plugin — registered in the Harness workspace through a `cordis.yml` patch (`id: harness-mcp-server`, `name: '@chushixixin/dsh-harness-mcp-server'`) and started with `corepack pnpm dsh web --patch …/cordis.yml`; inside Harness it runs a StreamableHTTP MCP server on `127.0.0.1:8090` and bridges `ctx.agents` / `ctx.agentPresets` / `ctx.tools` so an external MCP client drives real Harness agents

  • Five MCP tools over StreamableHTTP

    an MCP client pointed at `http://127.0.0.1:8090/mcp` (Hermes is the documented one)→`echo`, `harness_list_tools`, `agent_run`, `task_inbox` and `task_result`

    opens an HTTP listener on the configured `host` / `port` (the patch ships `127.0.0.1:8090`)the listener has no authentication by default — the README calls that unauthenticated remote code execution
  • Synchronous agent runs

    a task handed to `agent_run`→Harness creates an agent on the `standard` preset and returns the structured result of the run

    spawns a real Harness agent with the full toolset (`bash`, `fs`, `todo`, `web`) that operates on the cwd you send
  • Async task queue

    a task, memory context and cwd pushed through `task_inbox`→a queued task whose structured result `task_result` polls back

    the task waits inside the Harness process until a client polls it
  • Memory hand-off with the calling client

    memory context sent alongside each task→`changes`, `verification` and `leftovers` in every result, ready to be persisted back into the client's memory

  • Sandboxed bash and per-cwd session reuse

    bash commands the Harness agent decides to run→writes confined by the `workspace-write` sandbox, and an agent session reused per cwd across calls

    without `bubblewrap` on the host the sandbox refuses write commandseach new MCP session gets its own `McpServer` + transport, because an MCP `McpServer` connects to a single transport

04Who is it for? When not to use it?

Good for

  • Teams that keep the planning "brain" in a client like Hermes and want Harness's agent and full toolset as the "arms" doing the actual edits
  • Clients that have to stay responsive while Harness runs a long refactor or several unrelated tasks in parallel
  • Agents with persistent project memory that need every Harness run to start informed and end remembered
  • People who hit context limits in their main agent, or who want parallel execution of unrelated tasks

Not for

  • The listener is unauthenticated remote code execution: the README forbids binding it to `0.0.0.0` or exposing it to the internet/LAN before you add authentication, TLS and a reverse proxy.
  • Auth and a cwd allowlist are opt-in only — the `cordis.yml` example ships `authToken` and `workspaceRoots` commented out, so an unedited patch has neither.
  • Bash runs under the `workspace-write` sandbox, so without `bubblewrap` on the host the sandbox refuses write commands and file-changing tasks cannot land.

05Compatibility, maintenance and safety notes

  • The listener is unauthenticated remote code execution: the README forbids binding it to `0.0.0.0` or exposing it to the internet/LAN before you add authentication, TLS and a reverse proxy.
  • Auth and a cwd allowlist are opt-in only — the `cordis.yml` example ships `authToken` and `workspaceRoots` commented out, so an unedited patch has neither.
  • Bash runs under the `workspace-write` sandbox, so without `bubblewrap` on the host the sandbox refuses write commands and file-changing tasks cannot land.
  • The author positions it as a fallback tool rather than a daily driver, and every new MCP session gets its own `McpServer` + transport — one transport per server instance.
2026-08-132026-08-15v0.1.8

MIT · latest release v0.1.8 (2026-08-15T11:47:03Z) · early-stage plugin, 5 MCP tools

06Frequently asked questions

How do I load dsh-harness-mcp-server into a DeepSeek Harness workspace?

Install the package with `npm install @chushixixin/dsh-harness-mcp-server`, then reference it from the Harness workspace through the `cordis.yml` patch the README shows — the patch inserts plugin `harness-mcp-server` with `name: '@chushixixin/dsh-harness-mcp-server'`. There is no `dsh plugin add` step in the README: you start Harness with `corepack pnpm dsh web --patch ./packages/mcp/harness-mcp-server/cordis.yml`.

What actually runs inside the Harness process?

A StreamableHTTP MCP server on `127.0.0.1:8090` that bridges Harness's `ctx.agents`, `ctx.agentPresets` and `ctx.tools`. An external client calls `agent_run` or `task_inbox`, and Harness creates an agent on the `standard` preset with the full toolset (`bash`, `fs`, `todo`, `web`) to execute the task.

Is it safe to expose that MCP port?

No. The README states that the server exposes unauthenticated remote code execution and must not be bound to `0.0.0.0` or exposed to the internet/LAN without authentication, TLS and a reverse proxy first. `authToken` (Bearer) and `workspaceRoots` (cwd allowlist) are optional and commented out in the shipped patch.

Why does the author call it a fallback tool?

For everyday code edits he recommends driving your primary agent directly, and reaching for this only for context isolation (huge refactors that would blow the client's context) or parallel execution of unrelated tasks. The agent session is reused per cwd, which the README estimates at roughly 15–20× cheaper than one-shot `dsh headless`.

What do I need on the host before a write-heavy task?

`DEEPSEEK_API_KEY` must be exported for Harness, and `bubblewrap` must be installed — bash runs under the `workspace-write` sandbox and refuses write commands without it. Also note that each new MCP session gets its own `McpServer` and transport.

08Data and sources

  • Author-claimedgithub.comd174cea870aa…

    but it is a **Cordis app**, not something another agent can call.

  • Author-claimedgithub.comd174cea870aa…

    This plugin turns Harness inside-out: it starts a real **MCP server** (StreamableHTTP) *inside* Harness

  • Author-claimedgithub.comd174cea870aa…

    bridges Harness's core services — `ctx.agents`, `ctx.agentPresets`, `ctx.tools`

  • Author-claimedgithub.comd174cea870aa…

    corepack pnpm dsh web --patch ./packages/mcp/harness-mcp-server/cordis.yml

This page is generated from the project’s public documentation, repository metadata and a structured parse of DSH Plugins; last verified on 2026-11-16. Found an error? Submit a correction.

🏆

Best DeepSeek Harness Plugins

Twelve plugins worth installing first — picked from the whole catalog, across every category.

DSH Plugins is an independent community directory of DeepSeek Harness plugins. Not affiliated with or endorsed by DeepSeek. Third-party plugins are not security-audited — review the source before installing.

New DeepSeek Harness plugins, weekly. No spam.