appthin/dsh-mcp-manager-plus

English | 中文 | Changelog

An MCP server manager plugin for DeepSeek Harness: it adds an "MCP" page to the settings sidebar, where you can inspect, enable/disable, edit, restart, delete and add MCP servers.

Why this plugin exists:

• To manage MCP servers in DeepSeek Harness without ever opening a config file.
• Most MCP-related plugins on the plugin market target older DSH releases and simply do not work on current ones, so I had to write my own.

If this plugin stops working on a newer DSH release, please open an issue.

Screenshots

Settings → MCP: the server list, connection state and enable/disable toggles.

Expand a server: every tool the server currently registers with the model, with descriptions.

Add: paste an MCP configuration from any tool — the format is detected automatically (Codex's TOML and JSON-with-comments included).

Edit · form mode: edit field by field; switch to JSON mode at any time from the dialog header.

Edit · JSON mode: opens pre-filled with the server's current configuration; paste a whole replacement block, and renaming is just changing the name.

Installation

Requirements: DSH >= 0.1.6-alpha.1 and < 0.3.0 (tested on 0.1.6-alpha.1 and 0.2.0-rc.2). DSH itself needs Node.js >= 24.2.0 and Git >= 2.31.0.

Option 1: one-click install from the plugin market

If you have dshmarket installed, open Settings → Plugin market, search for dsh-mcp-manager-plus, and click install.

Option 2: command line

Pick any one of the three sources (--profile web installs into the web profile; substitute your own profile name as needed):

# npm package (available once published)
dsh plugin --profile web add dsh-mcp-manager-plus

# GitHub repository
dsh plugin --profile web add https://github.com/appthin/dsh-mcp-manager-plus.git

# Local directory (clone this repository first; installed as a link, so edits
# take effect immediately — handy for development)
dsh plugin --profile web add <the cloned repository directory>

Restart dsh (or wait for the profile hot-reload); the page appears under Settings → MCP.

Verify

Open dsh web and go to Settings → MCP (the chain-link icon in the sidebar). If the page does not show up, check in this order:

1. Switch to the latest Chrome/Edge — plugin bundles can fail to load on Chromium kernels older than 122;
2. Confirm it is installed into the profile you are actually running (dsh plugin --profile web ls lists the profile contents);
3. If the browser console shows errors, open an issue with the message attached.

Update & uninstall

dsh plugin --profile web ls                          # list installed plugins
dsh plugin --profile web add dsh-mcp-manager-plus    # update (re-adding installs the latest)
dsh plugin --profile web remove dsh-mcp-manager-plus # uninstall

With the plugin market installed, updates can also be done with one click on its page. Uninstalling keeps the MCP server rows this plugin wrote into the profile patch layer — that is your configuration data and will not be deleted.

Features

Capability	Description
List	Shows the MCP servers from the config file
Expand	Open a row to see every tool the server registers with the model, with descriptions
Enable / disable	Writes a disabled override into the patch layer; takes effect hot in ~1 s, no dsh restart
Edit	Two switchable modes: form (field by field) or JSON (one large textarea), see below
Restart	Destroys and re-initializes the server's connection in place (config unchanged); for recovering from dropped connections
Delete	Removes the server and all of its override rows from the config file
Add	Paste an MCP configuration from any mainstream tool; the format is detected, and multiple servers per paste are supported (see table below)
Bilingual	Chinese / English. Defaults to the harness UI language (falling back to the browser's language when unavailable); switch anytime from the page header — the choice is remembered, and host-side messages follow along

The two edit modes

When editing an existing server, the dialog header carries a segmented switch (with icons and a sliding indicator, radiogroup semantics, arrow-key navigable) to flip between modes at any time:

• Form mode: edit field by field (server name, transport, command, args, env, cwd, URL, headers).
• JSON mode: one large textarea that opens pre-filled with the server's current configuration (not an empty box or an example), so you can paste a whole replacement block — especially handy for configs copied from elsewhere. A collapsible format guide above the box lists the supported tools and their container keys, shared with the "Add" dialog.

Both modes write to the same row, therefore:

• Renaming in JSON mode renames the row: it keeps the original id and its enable/disable overrides instead of creating a duplicate;
• JSON mode supports every import format (Codex's TOML included), since format parsing is independent of add-vs-edit;
• A paste containing multiple servers is rejected with their names listed, avoiding the "which one actually got saved?" ambiguity;
• Switching back to form mode reads the row from disk, so half-edited JSON never leaks across modes.

JSON mode exists only for existing servers: saving JSON means "update the named row", which needs a row name. New servers go through "Add", which is already JSON mode.

The container-key table in the format guide (FORMAT_ROWS) must agree with the parser or the guide would lie — a test in test/client.test.mjs cross-checks both: every key in the table must actually parse, and all 14 tool names the parser supports must appear in the table.

Supported import formats

Paste as-is; the plugin recognizes container keys, field names and file syntax:

Tool	Container key	stdio	Remote	Notes
Claude Code · Cursor · Windsurf · Qoder · Cherry Studio · CodeBuddy · TRAE · ZCode	mcpServers	command/args	url/headers	the standard format
DeepSeek Harness (this plugin itself)	mcpServers	command/args	url/headers	native format
VS Code	servers	command/args	type:"http" + url	different outer key
Codex	mcp_servers	command/args	url	TOML, not JSON
OpenCode	mcp	type:"local", command is an array	type:"remote" + url	most divergent shape
Continue	mcpServers array	command/args	—	name lives inside each entry
Pi	mcpServers + settings	command/args	transport:"streamable-http"	settings is ignored

These variations are also handled:

• TOML input: Codex's config.toml can be pasted directly (including nested […​.env] tables and inline comments).
• JSON comments and trailing commas: docs for several tools carry //, /* */ and trailing commas — all parse fine.
• type:"sse" maps onto this plugin's streamable-http transport.
• serverUrl (Windsurf's remote spelling), http_headers (Codex) and command arrays (OpenCode) are all filed correctly.
• Unknown fields such as enabled, lifecycle or description never fail an import; they are just not written to the config.

The import reports the detected source, e.g. 已识别 Codex 配置；新增 xxx ("Codex configuration detected; added xxx"). A single failing entry is skipped and listed with its reason, so one bad block does not waste the whole paste.

Where the config lives

The plugin's single source of truth is the current profile's user patch layer:

$DSH_HOME/profiles/<profile>/cordis.patch.yml

Adding an MCP server appends one loader patch entry to that file:

- insert:
    - id: mcp-github
      name: '@deepseek-ai/dsh-mcp-client'
      config:
        serverName: github
        transport: stdio
        command: npx
        args: ["-y", "@modelcontextprotocol/server-github"]
        env:
          GITHUB_TOKEN: "…"

Disabling writes a sibling override row instead:

- id: mcp-github
  disabled: true

Because the profile runs with patchReload: live, the loader re-orchestrates the plugin tree within about a second of each save, so config changes are both persistent and hot-effective — the plugin never mutates the running loader tree; it only writes the file, then the browser briefly polls to observe the converged state.

Security model

• Loopback only. The API lives on the same loopback web server; non-loopback origins get a 403.
• Validate before writing. Every candidate file is re-parsed as a top-level loader patch array; a parse failure refuses the write, so one bad edit cannot break the next boot.
• Nothing outside its jurisdiction. The file is hand-written by users: comments, unrelated entries and unknown fields survive a rewrite untouched — only the one modified block is rewritten.
• Secrets stay local. Values whose keys look like KEY|PASSWORD|SECRET|TOKEN are masked on read, and unmasked again when the browser echoes them back; real secrets never appear on the page.
• Bundle-provided servers are read-only. Rows contributed by the bundle layer can be enabled/disabled but not edited or deleted.

Development

node test/run.mjs          # all seven suites (106 assertions)

Suite	Coverage
test/patch.test.mjs	patch-layer segmentation, parsing, validation, generation and byte-exact round-trips
test/import.test.mjs	config formats of 14 tools, the TOML subset, JSON comments and trailing commas
test/editjson.test.mjs	JSON-mode editing: in-place updates, renames keeping id and overrides, multi-server rejection
test/host.test.mjs	every host-side HTTP route, driving the real apply() through a fake Cordis context
test/client.test.mjs	whether the shell can load the browser bundle and register the settings page, mode switch included
test/compose.test.mjs	composing a real profile from the bundle layer, confirming this plugin's rows mount
test/live.test.mjs	read-only run against the real profile, confirming live configs read correctly with secrets masked

The tests use no test framework — just Node's built-in node:test — and reuse the js-yaml from the profile (the same parse path as the runtime).

Layout

File	Purpose
lib/index.js	host side: HTTP API, inventory projection, write operations
lib/import.js	import parsing: each tool's config format → this plugin's server shape
lib/patch.js	patch-layer read/write: segmentation, parsing, validation, generation
lib/client.js	browser side: the settings page UI (hand-written lazy-CJS bundle, no build step)
cordis.patch.yml	the bundle patch inserting this plugin into the profile's layer stack

Helper scripts

Script	Purpose
node tools/e2e-import.mjs	boots a real host and POSTs each of the 14 tool formats to /import, then validates the written patch rows
node tools/roundtrip-json.mjs	verifies the JSON editor's pre-filled config round-trips byte-for-byte (confirming unchanged content drops no fields)
node tools/check-client-bundle.mjs	loads the client bundle the way the shell does and checks the registration result and nav-icon CSS
node tools/check-segmented.mjs	checks the segmented switch's slider geometry (inset, half-width, offset) against the track assumptions
node tools/preview-edit-modes.mjs	renders both edit-dialog modes with the real components into edit-modes-preview.html
node tools/check-nav-icon.mjs	checks that the nav icon's mask payload is a drawable 16×16 template
node tools/gen-icon-preview.cjs	extracts icons from the dsh web bundle into the icon-options.html comparison page
node tools/dump-icon.cjs <IconName>	prints an icon's SVG path for hand-written overrides

License

MIT