dsh-mcp-status

English · Русский

A DeepSeek Harness plugin that puts an MCP button in the chat header. One click and you see which MCP servers are connected — and which of them answer right now, with the number of tools each one actually exposes.

Example servers, addresses and names in the screenshot are placeholders.

---

Contents

• Why this exists
• What you get
• How it works
• Wire format
• States
• Configuration
• Installation
• Standalone probe
• Troubleshooting
• Limitations
• License

Why this exists

DSH connects MCP servers once, at startup. When one of them is down, the symptoms are indirect: a tool that used to be in the model's toolset is simply not there, or a call fails with a transport error. The configuration file only tells you what should be connected, not what is.

An endpoint that holds a TCP port but never completes the MCP handshake looks "up" to any port checker. This plugin answers a stricter question: does the server complete initialize + tools/list right now, and how many tools does it return? That is the same thing the Harness itself needs in order to give the model those tools.

What you get

Element	Where	What it shows
Status pill	chat header, conversation.session.header.utilities	a dot: green when everything answers, red when something is dead
Server list	inside the popover	name, endpoint, state, response time, tool count
EDT block	inside the popover	which 1C:EDT projects are currently serving MCP, discovered from each project's .mcp.json
Refresh button	popover header	forces a probe; the timestamp changes when the new result lands

The answer is cached, so opening the panel is instant even though a full probe takes ~15 seconds.

How it works

Two halves in one package — the usual shape for a DSH plugin with a UI.

flowchart LR
    subgraph host["Host · Node"]
        A["apply(ctx)"] --> B["webServer.register<br/>GET /plugin/mcp-status"]
        A --> C["refresh()<br/>startup + every 5 min"]
        C --> D["spawn python<br/>mcp_status.py --json --out file"]
        D --> E[("cache<br/>data · fetchedAt")]
        E --> B
    end
    subgraph client["Client · browser"]
        F["McpStatusButton"] -->|"fetch, every 20 s"| B
        F --> G["popover<br/>servers + EDT"]
    end

Host half — index.js

Declares export const inject = ['webServer'], registers the route inside ctx.effect, and runs the probe at startup and then every five minutes. The probe is a child process: it writes JSON to a file, the plugin reads that file, caches it and serves it.

Client half — client.js

A classic script that registers itself through window.__ModuleLoader__.load and exports { inject: ['slots'], apply(ctx) }. It places a React component into the conversation.session.header.utilities slot — the same band as the locale and Markdown controls. The component polls the route on mount, then every 20 seconds; while the host reports refreshing: true it polls every 1.5 seconds, so the timestamp changes the moment the result is ready instead of half a minute later.

A handshake and not a port check

For every configured server the probe performs a real MCP handshake:

1. POST initialize with Accept: application/json, text/event-stream, following the 307 redirect some servers answer on /mcp before /mcp/.
2. Capture Mcp-Session-Id and send notifications/initialized — several FastMCP-based servers reject tools/list with "Invalid request parameters" without it.
3. POST tools/list with that session header; parse either a plain JSON body or an SSE stream (data: {...} blocks).
4. For stdio servers: spawn the configured command, speak newline-delimited JSON-RPC over stdin/stdout, with separate timeouts per stage — a router can take 12 seconds on tools/list because it rescans its port range first.

Why a Python child process

That logic already exists as a tested Python script (scripts/mcp_status.py, standard library only). Re-implementing the transport edge cases in JavaScript would create a second source of truth, so the host half simply runs the script and caches its JSON. The script ships inside the package — no external dependency.

One consequence worth knowing: the child's stdout is not captured through a pipe. The script writes its result to a file (--out), the plugin reads and deletes it. Pipes are fragile under sandboxed shells; a file handoff works everywhere.

Wire format

GET /plugin/mcp-status returns the cache envelope:

{
  "ok": true,
  "fetchedAt": 1793269934399,
  "durationMs": 14984,
  "refreshing": false,
  "error": null,
  "data": {
    "generatedAt": "2026-10-29T13:32:14",
    "config": "C:\\Users\\me\\.dsh\\profiles\\desktop\\cordis.patch.yml",
    "servers": [
      {
        "name": "edt-router",
        "transport": "stdio",
        "endpoint": "python edt-mcp-router.py",
        "state": "ok",
        "ms": 12293,
        "tools": 52,
        "detail": ""
      },
      {
        "name": "postgres",
        "transport": "http",
        "endpoint": "http://10.20.30.40:8000/mcp",
        "state": "dead",
        "ms": 2041,
        "tools": null,
        "detail": "порт закрыт (connection refused)"
      }
    ],
    "commentedOut": ["mcp-metadata"],
    "edt": [
      {
        "project": "My project",
        "url": "http://10.20.30.41:8765/mcp",
        "state": "ok",
        "tools": 50,
        "detail": ""
      }
    ],
    "orphanPorts": []
  }
}

Addresses above are examples and do not belong to any real network.

?refresh=1 starts a background probe and immediately returns the previous cache with refreshing: true. The UI keeps polling until refreshing flips back.

States

State	Meaning
ok	Handshake succeeded; tools is what the server returned.
warn	initialize succeeded, tools/list failed: alive but broken. detail carries the reason.
dead	Connection refused, timeout, HTTP error, or a non-MCP response.
off	The config row has disabled: true.
skipped	Not probed (for example an EDT project without a URL).

Configuration

All keys are optional and live in the loader row:

Key	Default	Meaning
pythonPath	the DSH runtime Python	Interpreter for the probe script.
scriptPath	scripts/mcp_status.py in the package	Your own copy if you change the probe.
intervalMs	300000 (5 min)	Background refresh period; minimum 15 s.

- insert:
    - id: mcp-status
      name: '@local/dsh-mcp-status'
      config:
        intervalMs: 60000

Installation

The package is a bundle: it ships a dsh.bundle.patch that inserts its own loader row.

1. Link it into the profile:

   cd ~\.dsh\profiles\desktop
   pnpm add "link:<absolute path to this package>"
   

2. Add the name to the bundle list in ~/.dsh/profiles/desktop/package.json:

   "dsh": {
     "profile": {
       "bundles": ["@deepseek-ai/dsh-base", "@deepseek-ai/dsh-web-app", "@local/dsh-mcp-status"]
     }
   }
   

3. Restart DSH. The host half activates with the profile; the client half appears on the next page load.

To remove it, drop the name from bundles and run pnpm install again.

[!NOTE] The name @local/dsh-mcp-status is arbitrary: a private package that is never published to npm. Rename it in package.json and in cordis.patch.yml, plus the profile's dependency key and bundle list.

Standalone probe

scripts/mcp_status.py runs without the UI:

python scripts/mcp_status.py                            # human-readable table
python scripts/mcp_status.py --json                     # machine-readable
python scripts/mcp_status.py --json --out result.json   # write to a file
python scripts/mcp_status.py --no-edt                   # skip the 1C:EDT block
python scripts/mcp_status.py --edt-roots 'D:\projects'  # EDT roots, when the router row carries none

Exit code 1 means at least one server is dead, 0 means all answered — handy for cron.

Troubleshooting

Symptom	Cause / fix
No button after install	The client half is registered on page load. Reload the page (F5), not just the app.
Route answers 404	The host half did not activate. Look for [mcp-status] warnings in the DSH stderr; check the loader row and that the package resolves from the profile.
cannot get property "webServer" without inject	Cordis forbids reading a service without declaring it. The plugin declares inject = ['webServer'].
"Checking…" for ~15 s	Expected: the stdio router rescans its port range on every tools/list. HTTP servers alone answer in ~200 ms.
dead for EDT projects	That EDT instance is not running. .mcp.json keeps the port from its last run — not an error.
Everything is dead	Nothing is reachable: check the MCP processes/containers, then press Refresh.

The package appends .activation.log next to itself when the module is imported and when apply throws. It is git-ignored.

Limitations

• The route is a plain webServer route, not an authenticated /api route. It exposes only names, endpoints and states on loopback, but if you bind DSH to a non-loopback address, this route is reachable too.
• The EDT block is 1C-specific; elsewhere it simply finds no projects.
• The client half receives no configuration (DSH creates client entries without it), so all tuning lives in the host row.

License

MIT