dsh-mcp-manager-ui

Click the floating button or the sidebar's "Toolbox" entry: the panel floats above the session, so changing configuration or reading skills never interrupts the conversation you are in. Two ability panes live in the same panel — MCP and skills.

The panel follows the DSH interface language: when DSH is set to English, every label, button, and status string in the panel is English too (the dictionary is EN_DICT in lib/client.js); with Chinese selected — or on a Host that ships no locale service — the original Chinese copy is shown. The language switch lives in DSH, not in the panel: Settings → General → Language. The panel deliberately has no switch of its own — the Host is the single source of truth for language, and two switches would inevitably disagree. Until you pick a language explicitly, the Host infers it from your browser language (an English browser means an English panel; pick Chinese once in Settings and the choice is remembered).

Screenshots

MCP market

47 services, each verified against first-party sources, across 10 categories. Search-first, with a "keyless only" filter and sorting. Every entry states its free tier honestly and what you must supply (key / connection details / arguments to edit after install); one click installs from the card, and already-configured entries are recognised and skipped.

Skills pane

See every installed skill's real scope and cross-scope shadowing, enable/disable, remove, read the body and diagnostics — and flip model auto-invocation with one switch, which edits disable-model-invocation in SKILL.md in place, touching only that one line.

Features

MCP

• Per-entry status, transport, connection settings and the registered tools; expand a tool for its full input JSON Schema (required/optional, types, enums, defaults)
• Enable, disable, reconnect, add, edit, remove; credential fields are masked by default — reveal the effective runtime value temporarily and copy it in one click
• Global + project scopes: global writes the Web profile (available to every project), project writes that project's .dsh/mcp.json (visible only to its sessions); a project can also block a global MCP
• Import configuration from other clients installed on this machine: Claude Code, Codex, OpenCode, pi, Claude Desktop, Cursor, Windsurf, VS Code, Gemini CLI, Roo Code and the cross-tool shared config — or paste JSON directly. Previewed before writing, with merge and replace modes
• Move servers between workspaces: copy this project's MCP servers into another registered workspace (done inside the Host, so credentials never pass through the browser), or export the whole project config as JSON
• Deferred tool loading (ToolSearch): a server's tools no longer have to stay resident in context every turn — the model first sees a name-only catalog and, once tool_search finds a tool (by keyword or select:<exact name>), its full parameter schema enters the tool list and becomes directly callable. Off by default: click the "Deferred / Resident" chip on a row to flip it (or use "Defer all" above the list); set DSH_MCP_TOOL_SEARCH=off to fall back to full injection everywhere. Its three limits (tools per search, catalog lines, discovered-set size) are adjustable under "Deferral settings" and stored as config.toolSearch in the profile patch

Skills

• Real scope and cross-scope shadowing: the same name in different scopes is a legal configuration; the panel lists each copy and marks which one wins
• Enable/disable (by writing the two invocation fields in the skill file, without touching the file itself), remove, read the body and diagnostics — the reason a skill "does not take effect" is right there in the panel
• Model auto-invocation switch: edits disable-model-invocation in SKILL.md; turning it off deletes that line (off → on round-trips byte-for-byte). The state always comes from the skill file itself — when it cannot be read the panel says "unknown" instead of guessing
• Located by its real position on disk (never by skill name): skills whose file or directory name differs from the frontmatter name can still be read, toggled and removed
• Market backed by skillhub.cn: publisher verification and security-scan findings are disclosed before install, followed by one explicit confirmation (validated server-side)

Shared

• Pick one or both entry points: the floating button (draggable, position remembered) and the sidebar "Toolbox" entry — at least one stays on
• Follows the DSH dark/light theme and adapts to narrow widths
• Update notice and one-click upgrade: at most one npm check per day, dismissible, never auto-updates; after an upgrade it explains what changed, without any network request

Install

# Recommended: from npm — `dsh plugin update` then upgrades within ^1.x
dsh plugin --profile web add dsh-toolbox-ui@^1.9.0

# Or pin a GitHub release tag (no automatic upgrades; re-add with a new tag)
dsh plugin --profile web add github:Imzl-zl/dsh-toolbox-ui#v1.9.0

Then restart dsh web. To uninstall (your existing MCP entries stay):

dsh plugin --profile web remove dsh-toolbox-ui

Do not add mcp-manager-ui to the profile's cordis.patch.yml, an agent preset, or an extra --patch file by hand — the plugin command already does it, and assembling it twice fails loudly.

Upgrade

dsh plugin --profile web update dsh-toolbox-ui

When a newer version exists the panel shows a dismissible banner; pressing Upgrade runs DSH's official pluginManager.

⚠️ 1.6.1 and earlier will not announce 1.6.2 or anything after it. Their update check read this repository's release feed, whose only latest entry was held by the legacy package dsh-mcp-manager-ui (frozen at 1.5.5) during that period — so it always answered "up to date". 1.6.2 fixed that by querying this package's latest on npm, and the fix can only ship inside a newer version: upgrade once by hand, then the banner works again.

dsh plugin --profile web update dsh-toolbox-ui

Still on the legacy package (≤1.5.5)? 1.5.5 keeps a permanent banner at the top of its panel — press "Migrate" and it installs this package and removes the old one. If the panel's Upgrade button reports ERR_PNPM_NO_MATCHING_VERSION, the legacy upgrade source has been retired (this repository's latest points at this package since v1.8.2); use the two commands instead:

dsh plugin --profile web remove dsh-mcp-manager-ui
dsh plugin --profile web add dsh-toolbox-ui

See Install & upgrade.

Quick start

Open the panel: the floating button at the bottom right, or "Toolbox" at the bottom of the sidebar. It floats above the session; Esc or "Close" hides it without touching the session.

Two scopes:

serverName is unique per scope: different projects, and global vs. project, may share a name and connect independently.

Adding MCP: use "Add MCP" for the form, or "Import MCP" to paste another client's configuration (previewed before writing). Common servers ship with presets.

FAQ

When does a change in a project take effect? The moment you save: new entries attach to live sessions, removed ones retract their tools, edits reload the connection in place — no need to reopen the session or restart the Host. Editing .dsh/mcp.json by hand is picked up when the next session mounts.

The connection follows the configuration, not the session. When the last session ends the connection is not torn down; it goes idle and the next session reuses it, so it is already there when you need it. Two knobs live on the server entry: idleTimeoutMs (how long an idle connection is kept, default 5 minutes, 0 = never reclaim) and readyTimeoutMs (the total budget for waiting until ready on mount, default 5s, 0 = do not wait). See Design & limits.

What do the project row's status strings mean? They separate two independent facts: whether a connection exists (已连接 · 当前无会话使用（连接保留，下次直接用） / 已连接 · N 个会话在用 / 连接中… / 连接失败 / 尚未建立连接（新会话自动挂载）) and how many sessions are using it — the old "waiting for a session to mount" wording merged both into one sentence and read like "not connected yet".

The panel says "disabled" but the model reports unknown tool "mcp__xxx__…". That MCP is disabled, so it registers no tools. Enable it in the panel.

With "Deferred" on, can the model still find those tools? Yes, just not by full visibility: the prompt carries a name-only catalog plus one line of guidance (no descriptions, no parameter schemas), and once tool_search finds a tool by keyword or select:<exact name>, its full schema enters the tool list and it becomes directly callable. Calling a not-yet-loaded tool from memory is refused, and the refusal hands back the exact query string to copy. Servers with only two or three tools save next to nothing (a guidance paragraph is still paid every turn); it pays off with many tools. To fall back to full injection everywhere: set DSH_MCP_TOOL_SEARCH=off. Note the "Deferred / Resident" chip on each row shows your configuration (that server's defer value) — click it to flip; it does not mean the mechanism is in effect in a given session right now.

How do I defer a single server? Every row carries a chip: "Resident" means that server's tools stay in context, "Deferred" means they load on demand. Click it to toggle — no need to open the edit form (global entries shown inside a project tab must be changed on the "Global" tab); "Defer all" at the top flips everything in the current scope at once. Servers with few tools still take effect when ticked; the panel just says the savings will be small.

Where do I tune the two limits? Panel → MCP → "Deferral settings" at the top: maxResults (tools returned per search, default 10) and maxDiscovered (discovered-set size, default 20). There is also a "Limit catalog size" switch (maxCatalogChars) that is off by default — off lists every tool in the catalog; turn it on only with very many tools and a need to trim context (it prefills 12500, a suggested starting point only). Saving writes them into config.toolSearch of the profile patch and takes effect in the current session immediately (no restart); values equal to the defaults are not written, so a future default change reaches you. You can also edit that config block by hand — then it applies after restarting dsh web. ⚠️ A misspelled key makes the plugin fail to load and reports the exact path (it is not silently ignored); fix it and restart.

How do I tell "mount failed", "connection failed" and "scope isolation failed" apart? Three different things: mount failed = it never attached during session mount; connection failed = mcp-client cannot connect or ran out of retries; scope isolation failed = the connection is fine but its tools are not in the shared scope layer. The detail pane gives the specific reason; see Design & limits.

Will it touch my other configuration? No. Global edits only touch the entries declared by @deepseek-ai/dsh-mcp-client and managed by this plugin, preserving other plugin entries, comments and !!js expressions. Uninstalling leaves your existing MCP entries and project .dsh/mcp.json alone.

Compatibility

Requires 0.2.x or newer: ToolSearch depends on the per-agent presentation mode (ctx.tools.modeFor(agent)) and the tools:sdk manual section, and neither exists on 0.1.x. No downgrade branches are kept for older hosts. The supported surface is the 0.2.0 and 0.2.1 lines: between them the host swapped loader generations (1.0.3 → 1.0.6), which changed startup-error propagation and what a dispose writes to disk. The plugin adapts to both explicitly, each with a standing test.

Development

npm test        # full suite, including host-level integration tests (real cordis / dsh-tools / dsh-scope / dsh-mcp-client + a real stdio MCP child process)
dsh web         # start the Web app for hands-on verification

Documentation

• Install & upgrade — three install flavours, release flow, troubleshooting
• Skills — scopes and shadowing, enable/disable, invocation policy, the skill market
• JSON import — field mapping and limits
• Project MCP lifecycle — mounting, the shared connection, scope alignment
• Design & limits — shared connection model, connection-state semantics, known limits, dependency details
• DeepSeek Harness plugin publishing guide

Friend link

License

MIT