dsh-mcp-panel

The MCP management console for the official DeepSeek Harness MCP client — add, edit, remove, and trial-call MCP servers from a settings page, with honest status, health diagnostics, and safe, reversible profile writes.

English · 简体中文 · Español · Português · हिन्दी

Architecture: official client = bridge, this plugin = console

@deepseek-ai/dsh-mcp-client is the only bridge: one plugin instance per MCP server, configured as a hand-written cordis.yml row, connecting the transport, syncing tools, and registering mcp__<server>__<tool> names. This plugin never replaces it — it is the experience layer on top:

                    ┌────────────────────────────────────────────┐
 profile            │  cordis.yml / cordis.patch.yml             │
 composition        │   - id: mcp-github                          │
 (one row per       │     name: '@deepseek-ai/dsh-mcp-client'     │
  server, hand-     │     config: { serverName, transport, … }    │
  written)          │   - id: mcp-panel                           │
                    │     name: dsh-mcp-panel   ◄── this plugin   │
                    └───────────────┬────────────────────────────┘
                                    │
        ┌───────────────────────────┴───────────────────────────┐
        │                                                        │
   ┌────▼──────────────┐        ┌───────────────────────────┐    │
   │ @deepseek-ai/dsh- │        │ dsh-mcp-panel (console)   │    │
   │ mcp-client        │        │                           │    │
   │ • transport       │        │ • /mcp command            │    │
   │ • tool sync       │        │ • Settings → Plugins →    │    │
   │ • mcp__* tools    │◄──────►│   MCP tab: CRUD, trial    │    │
   │ • mcp/status seam │ status │ • health diagnostics      │    │
   └───────────────────┘        │ • probes, capabilities    │    │
                                └───────────────────────────┘    │

The console reads the client through its shipped mcp/status observability seam (event + mcpStatus query service), the tool registry, and the loader; it writes only the profile's patch layer — append-only, approval-gated, always backed up. Transport, OAuth, and protocol stay untouched.

Console vs. hand-written cordis.yml

The console's output IS cordis.patch.yml vocabulary — the same lines you would write by hand, generated, previewed, and applied safely.

What you get

Quick start

# git channel (builds via the package's prepare script)
dsh plugin --profile web add github:PerryLink/dsh-mcp-panel#v0.4.0
# npm channel (published tarball, no build approval needed)
dsh plugin --profile web add dsh-mcp-panel@0.4.0

Then restart (or let the web surface hot-reload cordis.patch.yml) and open Settings → Plugins → MCP, or run:

/mcp
/mcp everything tools
/mcp everything health
/mcp everything call echo '{"message": "hi"}'

Manual install: put dsh-mcp-panel into the profile's node_modules (or the shared $DSH_HOME/profiles/node_modules fallback) and add the row to cordis.patch.yml:

- insert:
    - id: mcp-panel
      name: dsh-mcp-panel
      config:
        probeEnabled: true

Uninstall

1. Remove the mcp-panel row from cordis.patch.yml (the web surface hot-reloads it; other surfaces restart).
2. Delete the package from the profile's node_modules (or the shared profiles/node_modules fallback).
3. Verify with dsh web --dump-config that no mcp-panel row remains.

Honest by contract

• The bridge stays the bridge. No transport, OAuth, or protocol changes; one mcp-client row per server, exactly as hand-written.
• No fake status. Connection fields without upstream observations read unknown / — with statusSource: 'derived'; exit codes / stderr tails are never invented.
• Sanitized display. URL query credentials, userinfo passwords, header values, bearer tokens, and JWTs are redacted before rendering; configured headers never enter any snapshot; env/header values never leave the host (the editor sees keys only).
• Writes are append-only, approval-gated, and backed up. The console never rewrites cordis.patch.yml: it appends generated operations. When an approval service exists and the caller's session has a live agent inside an open turn, the write asks ctx.approval (only allowed-once proceeds); otherwise the explicit interactive confirmation is the approval channel. writeEnabled: false is a hard kill switch.
• No prompt injection. The panel registers no prompt sections; its only model-facing text is the two tool/command descriptions, in the official client's minimal style.

Config

Resources & Prompts

The official client documents "Tools are the only bridged MCP capability" — Resources and Prompts are deferred. The console feature-detects a proposed upstream catalog seam and will show read-only lists the day it ships; until then the capabilities board marks both pending upstream support (see the harness docs/upstream-proposal.md addendum for the follow-up proposals).

Development

pnpm run typecheck && pnpm run typecheck:ci && pnpm test && pnpm run build && pnpm run verify:self-contained && pnpm run verify:artifacts && pnpm pack

• src/patch.ts — validation, keep-semantics merge, YAML fragment rendering (pure).
• src/write.ts — backup + append + retention (the only file-write module).
• src/trial.ts — official-pipeline trial calls via ctx.tools.execute().
• src/diagnostics.ts — error-pattern → suggestion mapping (pure).
• src/client/ — the settings console (server editor, trial console, diagnostics).
• scripts/verify-headless.mjs boots the real web profile and prints exact /mcp output.

Releases: node scripts/release.mjs <x.y.z> runs the full gate, commits, and tags v<x.y.z> locally (never pushes).

Contributors

Thanks to everyone who reported issues, reviewed, or contributed code — in particular xiaoyuyu6420, who diagnosed the missing client devDependencies behind clean-checkout build failures (PR #5).

License

Apache-2.0 — see LICENSE.