dsh-mcp-manager

dsh-mcp-manager

简体中文 | English

Local 0.13.3 unreleased candidate: see the release notes for cumulative changes, provenance and verification limits. It supersedes 0.13.2.

MCP server manager for DeepSeek Harness (DSH) — a Settings → MCP page where you add MCP servers once (remote HTTP or local stdio process), authenticate HTTP servers with OAuth in the browser, and expose their tools either directly or through a compact on-demand broker.

The built-in @deepseek-ai/dsh-mcp-client only accepts a static headers config — it has no OAuth support and no local stdio transport. This plugin fills that gap:

• OAuth (authorization code + PKCE) with spec-compliant discovery, RFC 7591 dynamic client registration, refresh_token rotation, and auto-reconnect across restarts — one browser login, then it keeps working.
• Client application type follows MCP SEP-837: HTTP loopback callbacks (localhost, IPv4 loopback, or [::1]) register as native; other callbacks register as web. This avoids OIDC providers defaulting a local GUI to web and rejecting its redirect. Registration failures show the provider's OAuth error code when available; the Logs panel retains the detailed reason. Providers may still restrict which clients or redirects they accept.
• Callback origin accepts direct loopback Host values (including [::1]:port). Configure

publicOrigins for a reverse proxy: one configured origin is the default callback even without forwarded headers; with multiple origins, a complete X-Forwarded-Proto/X-Forwarded-Host pair must match the allowlist. A foreign direct Host is rejected when no default is configured. Malformed forwarded host values are rejected. Callback messages are HTML-escaped.

• Discovery follows RFC 9728 → RFC 8414: the MCP endpoint's 401 WWW-Authenticate header names its protected-resource metadata, that document names the authorization server, and the authorization server's metadata is read from the RFC 8414 path-inserted URL (e.g. https://github.com/.well-known/oauth-authorization-server/login/oauth). Only when nothing is discoverable does it fall back to guessing ${origin}/oauth/*, and the attempt log says so.
• RFC 8707 `resource` taken from the protected-resource document and sent on authorize, token

exchange and refresh. A server that binds tokens to an audience (Kontrcheck's connector does) rejects every step without it as invalid_target. Providers that publish no resource keep the previous behaviour exactly: nothing is invented.

• Manual client credentials for authorization servers that do not support dynamic registration (GitHub's does not): set Client ID to an app you registered with the provider, optionally a Client secret environment variable name (sent as client_secret; the secret itself is never persisted), and a Scope string.
• Static Bearer token mode for servers without OAuth — stored as an environment-variable name (Codex-style tokenEnv), never as plaintext in the config.
• No auth mode for servers that need no credentials at all (e.g. a local http://127.0.0.1:9316/mcp) — the plugin sends no Authorization header and connects on save.
• Custom HTTP headers (headers for direct values, headerEnv for values read from environment variables) — matches Codex's http_headers / env_http_headers.
• stdio local processes: run npx / uvx / python etc. directly; the plugin speaks JSON-RPC over the child's stdin/stdout (spawns the process, reconnects, and reaps it on exit) — no remote server or auth required. Windows .cmd shims (e.g. npx.cmd) are resolved through cmd.exe.
• HTTP-with-SSE servers: the legacy SSE transport (a GET event stream that advertises a session endpoint, with replies delivered asynchronously on that stream) is detected automatically and used when a server rejects a plain POST. Streamable HTTP servers are unaffected.
• Edit-in-place: rename a server, switch stdio ↔ HTTP, or change auth/headers without deleting and re-adding it.
• In-GUI diagnostics: a failed server shows its reason on the collapsed card (no need to expand it), and the expanded card has a Logs panel — an on-demand, redacted trace of every discovery URL and HTTP status, client registration, authorization, token exchange/refresh, initialize, tools/list, and the stderr tail of a stdio child that exited. Secrets (bearer values, code/state, client_secret, tokens) are stripped when the entry is written, never when it is rendered.
• Tool registration with the same mcp____ naming convention as the built-in client, including strict-schema sanitization for the DSH tool registry and isConcurrencySafe marking.
• Workspace isolation: declare per-project servers in /.dsh/dshmm/mcp.json — their tools register only into that workspace's sessions, and you can mask specific global servers per workspace.
• Opt-in on-demand broker: keep the model-facing MCP surface fixed at mcp_search_tools, mcp_describe_tool, and mcp_execute_tool instead of sending every mcp__* schema on every Native-mode request. It is disabled by default. mcp_search_tools is a zero-dependency lexical ranker (BM25 + CJK/alias/fuzzy, plus a catalog-listing fallback).
• Stable tool refresh: notifications/tools/list_changed refreshes only added, removed, or schema-changed registrations for both stdio and Streamable HTTP servers.

The MCP UI follows DSH's language preference in Settings → General → Language through the mcp locale namespace. DSH owns browser fallback, preference persistence, and live language updates; the plugin has no separate language selector or language API. Labels, forms, status badges, and confirmations are translated; server-supplied diagnostics remain in their original language. Legacy language values in ~/.dsh/mcp-manager.json are ignored and left untouched.

Version 0.11.0 requires @deepseek-ai/dsh-client-locale and the slot system's locale t prop. The compatibility runs listed below predate this integration and do not verify live locale switching on every listed DSH release. Recheck Settings → MCP while switching DSH's language before deploying to an older web profile.

Requirements

• DeepSeek Harness, web profile for the Settings → MCP page (npx @deepseek-ai/dsh web). On a profile without a GUI webserver (headless/tui) the plugin still registers MCP tools and connects servers — only the settings page is missing.
• Node.js ^22.19 or >=24; pnpm on your PATH
• Verified DSH releases are declared one by one in package.json → dsh.compatibility.dshReleases (0.1.2-rc.1, 0.1.5-alpha.1, 0.1.5-alpha.2, 0.1.5-rc.1 are compatible; supported range >=0.1.2-rc.1 __ is also visible to the agent), then delete the profile — never against a real ~/.dsh.
• Windows 10/11: stdio commands are launched via cmd.exe so .cmd shims (npx, uvx) resolve correctly

Install

npx -p @deepseek-ai/dsh dsh plugin --profile web add github:hyqhyq3/dsh-mcp-manager

Then restart dsh --profile web and refresh the page. The package declares a dsh.bundle.patch, so the plugin activates automatically — no manual cordis.patch.yml editing.

The MCP server's OAuth provider must allow a loopback redirect (http://127.0.0.1:/mcp-manager/callback/), which is where the DSH GUI webserver receives the code. The origin is derived from your browser's own address, so any host/port the GUI is served on works.

Usage

1. Open Settings → MCP in the DSH web UI.
2. ＋ Add MCP server (and later 编辑 / Edit to change it):

• Scope (作用域): user — a global server available in every workspace; or workspace — a server bound to one workspace (its config lives in that workspace's .dsh/dshmm/mcp.json). Pick the workspace from the second dropdown.
• HTTP: name (becomes the mcp____* prefix), URL, auth mode (OAuth, static token, or no auth), and optional headers (headers direct values, headerEnv values read from env vars). In OAuth mode you can additionally pre-register the client: Client ID, Client secret environment variable, and Scope (all optional).
• stdio: name, command (e.g. npx), args (one per row), env vars (key/value rows), and optional working directory.

1. OAuth servers: click 去认证 (Authenticate) → the browser opens the server's login page → after consent you are redirected back and the tools are registered immediately. If the authorization server advertises a registration_endpoint the plugin registers itself (RFC 7591) and you can leave Client ID empty. If it does not — GitHub's https://github.com/login/oauth is the common case — register an OAuth App with the provider, paste its Client ID (and the name of the environment variable holding its secret, if it is a confidential client), set Scope to what the MCP endpoint needs (e.g. repo read:org), and whitelist http://127.0.0.1:/mcp-manager/callback/ as the redirect URI there. Without a Client ID the attempt fails with an explanation rather than a bare HTTP 404, and the Logs panel lists every URL that was tried.
2. Static-token servers: enter the name of an environment variable that holds the token (e.g. MCP_BEARER_TOKEN) — the token itself is never written to disk; stdio servers spawn and connect immediately on save.
3. No-auth servers: pick No auth (server needs no credentials) — the plugin sends no Authorization header and connects on save. Use it for an endpoint that authenticates nothing (for example a local http://127.0.0.1:9316/mcp).
4. Optional: turn on On-demand MCP tool calls at the top of the page. The setting is profile-wide, persists across restarts, and affects existing sessions on their next request.

Status badges: connected (N tools) / needs-auth / authorizing / error / disabled. Buttons: authenticate, edit, enable/disable (switch), delete. A non-empty error is also shown as a one-line summary under the server name while the card is collapsed, and the expanded card offers 日志 / Logs. Disable unregisters that server's tools and drops its connection (config and OAuth tokens persist); Enable reconnects without re-authenticating. Disabled servers stay dormant across restarts. The toggle is global: it affects every session in this profile. State persists at ~/.dsh/mcp-manager.json (server configs + OAuth client registrations + tokens; static tokens are referenced by env-var name, not stored).

Tool-call timeouts

Since v0.14.0, the add/edit form has an optional Tool-call timeout (ms) for both HTTP and stdio servers, in user and workspace scope. For example, 600000 allows a simulation tool to run for ten minutes. Clear the input to inherit the default. The same field works in /.dsh/dshmm/mcp.json:

{
  "mcpServers": {
    "simulation": {
      "type": "http",
      "url": "http://127.0.0.1:9316/mcp",
      "authMode": "none",
      "timeout": 600000
    }
  }
}

Resolution order: server timeout, then the plugin row's config.toolCallTimeoutMs, then 60000 milliseconds. Values must be finite and at least 1000; numeric strings are accepted and fractions are floored. Invalid, empty, or missing values are treated as unset. PUT clears the server override when timeout is omitted or invalid. Disabling a server preserves its override.

The override applies only to tools/call over stdio, Streamable HTTP, and legacy HTTP-with-SSE. Initialization, tool listing, and notifications keep their existing 60-second defaults. Timeout errors include the elapsed budget; a client timeout does not guarantee that a remote server has stopped its work. A surrounding Code/PTC runtime may impose its own deadline.

What the agent sees

With on-demand mode off (the default), every connected server's tools appear as first-class tools, e.g. for a server named odin:

mcp__odin__search_tools     mcp__odin__describe_tool
mcp__odin__execute_tool     mcp__odin__list_tool_scopes

Tool results are projected back as native DSH content blocks; MCP isError results surface through the registry's error path. Image blocks ({ type: 'image', data, mimeType }) are never forwarded raw — a raw MCP image block has no attachment, so it would crash the session on the next turn. Rendering is text-first: when the routed model declares image input, the host stores the image in DSH's durable attachment store and the block becomes a real { type: 'image', attachment }; in every other case (no attachment store, model without image input, non-canonical base64, or a media type outside PNG/JPEG/WebP/GIF) the image degrades to an [image unavailable: …] text placeholder. Text blocks are preserved in order, and the same sanitizing applies to results returned through mcp_execute_tool.

With on-demand mode on, a Native-mode agent sees only these three MCP broker tools:

• mcp_search_tools({ query?, server?, limit? }) is a pure-lexical ranked search (no embedding model, no network): NFKC normalization, camelCase/CJK tokenization with a small English stemmer, a built-in bilingual alias table (e.g. 查日志 → log/logs), BM25 over the tool/server/description fields, and a bounded edit-distance fallback for typos. Omit query to list the catalog (optionally server-filtered). It returns at most limit matches (default 10, clamped to 1-50) plus total, the number of results before truncation.
• mcp_describe_tool({ name }) returns the exact registered description and input schema for one tool visible in that session.
• mcp_execute_tool({ name, arguments }) executes any currently visible MCP tool through the normal DSH tool pipeline. Calling describe first is recommended but not required.

Raw mcp__* names are removed from the model request and direct calls to them are denied; only the nested dispatch owned by mcp_execute_tool is allowed. Search, describe, and execute all resolve the calling agent's live registry view, so workspace isolation and exclude masks remain effective.

Code/PTC/both agents retain their full session-visible MCP catalog, even with this switch on. The broker remains callable but does not shrink the generated SDK. DSH 0.2.0-rc.2 names its modes native/ptc/both; 0.2.1-rc.2 supports native/ptc. See issue #18 investigation and isolated verification for the checked interfaces, safe fallbacks, context-cost measurements and test commands. The local 0.13.3 candidate also keeps this conservative fallback when a custom prompt empties tools:sdk but still exposes run_code.

Workspace isolation

Global servers (added in Settings → MCP) are visible in every workspace. Use the workspace dropdown at the top of the Settings → MCP page to switch between "global" and a specific workspace; when a workspace is selected you see both its own servers and the global servers (with a 隐藏 / Hide toggle to mask each global server). Workspace servers live in /.dsh/dshmm/mcp.json (Claude/Codex-style):

{
  "mcpServers": {
    "filesystem": { "type": "stdio", "command": "npx", "args": ["-y", "@modelcontextprotocol/server-filesystem", "."] },
    "unity-mcp": { "type": "http", "url": "http://localhost:8090/", "authMode": "static", "tokenEnv": "UNITY_MCP_TOKEN" },
    "local-mcp": { "type": "http", "url": "http://127.0.0.1:9316/mcp", "authMode": "none" }
  },
  "exclude": ["github"]
}

• Workspace servers can be added / edited / deleted in the UI for workspaces registered in DSH (the ＋ button while a workspace is selected writes to that workspace's mcp.json). Hand-editing the file also works — creating or changing it is hot-reloaded. Invalid JSON is shown as an error while the last valid live configuration stays active.
• type defaults to http; a stdio server's cwd defaults to the workspace root. headers / headerEnv / env / args follow the same shapes as the Settings form.
• A workspace server's tools register only into sessions whose working directory resolves to that workspace; another workspace's agents never see them. Global servers stay visible everywhere unless masked.
• exclude lists global servers to hide in this workspace (their tools are masked through the tool registry's per-agent restriction). Toggle it via the 隐藏 / Hide checkbox on each global server in the workspace view.
• serverName must be unique across global + all workspace sources; a later duplicate is flagged as a conflict and skipped (shown in the UI). A conflict is re-derived when the blocking global server is deleted or renamed, so the workspace entry connects on its own instead of waiting for a mcp.json edit.
• Config is re-read on each new session and hot-reloaded via a file watcher.
• Workspace connection failures retry automatically after 3s, 6s, 12s, up to 60s; a successful connection resets the delay. SSE session loss reconnects immediately and refreshes tools for existing agents. Credential-gated servers remain needs-auth. Editing/deleting a server, closing its last session, or unloading the plugin cancels pending retries.
• Workspace Logs work for every row, not just loaded ones: ids are deterministic (ws-), so a row listed from disk keeps the same id after it is loaded, and GET /mcp-manager/api/servers/:id/logs returns 200 for loaded, conflict and not-yet-loaded rows alike. POST .../connect reconnects loaded rows, answers 409 for a conflict row (server name is taken by another server), and 404 when nothing is loaded for that id. Logs retain redacted failure causes (such as ECONNREFUSED) across config rescans; deleting the server clears them.

Recovery regression tests run with node --test test/workspace-reconnect.test.js. For an opt-in real Loader smoke, install @deepseek-ai/dsh@0.2.1-rc.2 into a disposable npm prefix, then run node scripts/verify-real-dsh.mjs . The script creates and deletes its own temporary profile and workspace, uses loopback MCP only, and submits no model request. Verified on macOS / Node 24.12.0; Windows remains unverified.

• Workspace servers support stdio, HTTP static-token (tokenEnv), HTTP OAuth (the same discovery + PKCE + dynamic client registration flow as global servers, including the optional clientId / clientSecretEnv / scope fields in the hand-written config), and HTTP no-auth (authMode: "none", no Authorization header). Workspace OAuth tokens persist in ~/.dsh/mcp-manager.json (never in the declarative mcp.json); each workspace server row exposes a 去认证 / Authenticate button for OAuth servers.

How it works

In v0.13.4, pre-registered OAuth clients reuse the configured Client ID for both token exchange and refresh; client secrets are read from their configured environment variable for each grant. Global connections discard handshakes and failures that finish after disable, delete, edit, or plugin unload, so stale attempts cannot restore tools or clear a newer connection.

Piece · Mechanism Settings page · Client half registers a settings.section slot entry (MCP tab) OAuth flow · Host half discovers the authorization server (RFC 9728 protected-resource metadata → RFC 8414 authorization-server metadata), then does dynamic client registration (or uses a manually configured Client ID) + PKCE; the redirect lands on a route mounted on the DSH GUI webserver itself OAuth discovery · The MCP endpoint's 401 WWW-Authenticate is parsed for resource_metadata, which is cached on the server; resource metadata names the authorization server, whose metadata is fetched from the RFC 8414 path-inserted well-known URL (plus an OpenID-configuration and origin-level probe). Guessed ${origin}/oauth/* endpoints are a last resort, tagged as a fallback Diagnostics · A per-server in-memory ring buffer (last 20 entries) records discovery/register/authorize/token/initialize/tools-list/stdio phases with URL + HTTP status, exposed at GET /mcp-manager/api/servers/:id/logs; entries are redacted on write and never persisted (a log file would be a new secret-at-rest surface) Token storage · ~/.dsh/mcp-manager.json; OAuth tokens refreshed automatically on 401. Static tokens are read from the environment variable named by tokenEnv, and an OAuth client_secret is read from the environment variable named by clientSecretEnv — neither value is ever persisted. No-auth servers store no credential and send no Authorization header Refresh safety · Refresh is single-flight per process and serialized across DSH processes that share the state file: a shared lock covers refresh and all configuration writes, with FIFO ownership within each process so queued saves cannot overtake each other. A re-read adopts the latest disk pair even when its access token has expired, and changed-field merging preserves other processes’ server/workspace tokens. Stale locks are reclaimed only for dead or incomplete owners; live owners are never evicted by age. Reclamation is itself exclusive; an orphaned .refresh.lock.reclaim guard after a crash requires operator inspection/removal before stale recovery resumes (writes time out rather than steal a lock). The state file is replaced atomically (0600, fsync + rename). If a rotated pair cannot be persisted, that process keeps its access token but stops further refreshes until successful reauthorization; repair the state path before reauthorizing. An invalid_grant marks the server needs-auth and stops replaying the spent token. /api/servers reports oauth.expiresAt, needsReauth, lastRefreshAt, lastRefreshError, lastToolOkAt — timestamps and error codes only, never token material Legacy state · On load, servers without an id are assigned one (and persisted), and legacy [{ name, value }] env/header lists are normalized to maps — without this, id-addressed APIs 404 and array env values are dropped silently MCP transport (HTTP) · Streamable HTTP (JSON-RPC over POST, Mcp-Session-Id, SSE or JSON responses) with automatic fallback to the legacy HTTP-with-SSE session transport — a GET stream advertises the POST endpoint and delivers replies asynchronously; custom headers/headerEnv merged into every request Reconnect · Startup uses the existing immediate connection path; disabling or deleting a server does not queue a second delayed restore. A failed connect retries with exponential backoff (3s → 6s → 12s … capped at 60s, reset on success); an SSE session whose stream drops immediately reconnects, and its in-flight calls fail fast instead of blocking until timeout MCP transport (stdio) · child_process.spawn a local command, JSON-RPC over stdin/stdout (newline-delimited); reconnect reaps the old process first. On Windows it spawns through cmd.exe so .cmd shims resolve Tool schema · Server JSON Schemas are sanitized to the registry's supported raw subset (unsupported vocabulary degrades to unconstrained) Image results · output.render is text-first; execute(args, exec) stages a projection and finalizeContent installs durable DSH attachments only when the routed model declares image input — every other case degrades to a text placeholder | On-demand broker | A profile setting installs three broker tools, filters raw mcp__* schemas after