dsh-mcp-host

dsh plugin --profile web add dsh-mcp-host

dsh-mcp-host

Expose any stdio MCP server as native DeepSeek Harness tools. Host-only plugin, no browser UI, no npm dependencies.

Built for one job: let a DSH agent drive a browser through the official Microsoft Playwright MCP server, attached to the Edge/Chrome you are already using.

0.4.0 — the browser tools unload when the browsing turn ends (and right after edge_close); later requests carry only edge_tools. The MCP process and the open browser session stay alive, so reloading is one cheap call. See [Unload after use](#unload-after-use-default).

What it does

1. On load, spawns the MCP server once (default npx --yes @playwright/mcp@latest --cdp-endpoint=msedge) and calls initialize + tools/list.

2. Registers one bootstrap tool, edge_tools (name configurable). Its description is deliberately short and states when to call it — and when not to.

3. When the model calls edge_tools, the bridge registers the real edge_* tools for that Agent only, hides the bootstrap entry, and forwards every call to MCP tools/call. Screenshots are written to a temp file and returned as a path, not as base64.

4. Unload after use: when that turn ends (session/event → turn/end), the per-Agent tools are removed and edge_tools becomes visible again, so every later request is back to one small schema. edge_close releases immediately, and an idle countdown (MCP_BRIDGE_RELEASE_MS, default 3 min) is the backstop. The MCP process and the open browser session stay alive, so reloading is one cheap call. On agent/disposed (and on plugin dispose) the same teardown runs.

Tools are renamed: browser_navigate becomes edge_navigate (prefix configurable).

Lazy loading (default)

The 12-tool default set is not in the prompt until a browsing task needs it. A session that never touches a browser pays for one small schema (edge_tools), not twelve.

• Set MCP_BRIDGE_EAGER=1 to restore the old behaviour (register every tool at boot).
• Activation is per Agent: the parent session and each subagent must call edge_tools independently.
• After activation the bootstrap entry is hidden for that Agent (tools.restrict({ deny })), so the prompt goes back to a single, smaller surface.

Unload after use (default)

Loading is temporary, so a session that browsed once does not carry twelve schemas forever:

• Primary trigger: the end of the turn. As soon as the session emits turn/end, that Agent's tools are disposed and edge_tools appears again — the next request only pays for the small bootstrap schema. The next browsing turn reloads them with one call.
• edge_close: immediate unload (the page is closed, so the burst is over).
• Idle backstop: if no turn/end ever arrives (non-interactive hosts, long-lived agents), the tools also unload after MCP_BRIDGE_RELEASE_MS without a browser call. 0 disables the backstop.
• The MCP server process is not killed and the browser stays open — only the prompt surface shrinks, and reloading keeps tabs, logins and history.
• MCP_BRIDGE_RELEASE_ON_TURN_END=0 (with MCP_BRIDGE_RELEASE_MS=0) restores the 0.2.0 behaviour: tools stay for the whole Agent lifetime.

Default tool set (12 of Playwright's 25)

edge_navigate, edge_navigate_back, edge_snapshot, edge_click, edge_type, edge_fill_form, edge_press_key, edge_tabs, edge_take_screenshot, edge_evaluate, edge_wait_for, edge_close.

Set MCP_BRIDGE_ALL=1 to expose everything the server offers.

Screenshots: image, path, or both

Playwright MCP ≥ 0.0.83 saves every screenshot to disk and returns text containing the path ([Screenshot of viewport](.playwright-mcp\page-....png)), plus an image part unless told otherwise. The path is relative to the MCP server's working directory (on this machine the profile dir) unless MCP_BRIDGE_COMMAND sets an absolute --output-dir.

This plugin never forwards that image part. normalizeResult writes images to a temp file and every tool renders JSON text, so the model sees a path (~26 tokens), not an image. A screenshot has always been cheap here.

So MCP_BRIDGE_IMAGES does not save context in DSH. What it buys:

• MCP_BRIDGE_IMAGES=omit — the upstream stops sending the image at all: ~12 K fewer base64 characters over stdio per screenshot, and one less duplicate temp file. Pick it when the agent never looks at the picture. Caveat: with omit the only path is the relative one inside the text, because the plugin's absolute temp copy (images[]) only exists in the default mode — an agent that wants to open screenshots should keep allow.
• MCP_BRIDGE_IMAGES=only — keep the image, drop the accompanying text.
• Appended only to Playwright MCP command lines (@playwright/mcp or playwright-mcp); other servers are left untouched, and an explicit --image-responses inside MCP_BRIDGE_COMMAND always wins.
• Default is unchanged (allow).

Context only suffers where a client renders MCP image content (Claude Code, Cursor, …): there a 1280×720 capture is ~1.2k tokens. Those clients do not use this plugin — set --image-responses=omit in their own MCP config instead.

Output hygiene

Playwright MCP writes every screenshot and accessibility snapshot into its output directory (--output-dir, default <cwd>/.playwright-mcp) and never rotates it, so a long browsing life leaves hundreds of files behind. The bridge prunes once at plugin load and after every tool call: it keeps the newest MCP_BRIDGE_KEEP_OUTPUT files in that directory and in its own temp image directory (%TEMP%\dsh-mcp-bridge), deleting older .png/.jpg/.jpeg/.webp/.yml/.yaml/.log files. Anything else in those directories is left alone, and the screenshot just taken is always the newest, so reading it still works. 0 deletes all managed files; a missing directory or an unparsable value is a no-op.

Install

From npm (published as dsh-mcp-host):

dsh plugin --profile web 

…(truncated)