@adenxie/dsh-vercel
DeepSeek Harness Vercel MCP 连接，支持 OAuth、设置面板和按需使用的官方 Vercel 技能
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:AdenXie/dsh-vercel#9af0ee93de7729d0f3a3756febece4bd5729b151Vercel 以 @adenxie/dsh-vercel 发布，当前版本为 0.3.1。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
Vercel projects, deployments, and platform tools for DeepSeek Harness.
Connect with OAuth and load Vercel skill guidance when you need it.
English · 简体中文
@adenxie/dsh-vercel0.3.1. I have verified the plugin in DSH Desktop and an isolated Host profile. See the verification record for the checks and their limits.
Initialize and fully quit DSH Desktop, then install the npm package with the bundled CLI:
$DshCli = Join-Path $env:LOCALAPPDATA 'Programs\DeepSeek Harness\resources\runtime\cli\bin\dsh.cmd'
& $DshCli plugin --profile desktop add '@adenxie/dsh-vercel@0.3.1'
Restart Desktop afterward. For a Web profile, use dsh plugin --profile web add @adenxie/dsh-vercel@0.3.1.
Build the local checkout and run npm pack to produce adenxie-dsh-vercel-0.3.1.tgz. I recommend this packaged installation for regular use. Initialize and fully quit Desktop, then run PowerShell in the archive directory:
$DshCli = Join-Path $env:LOCALAPPDATA 'Programs\DeepSeek Harness\resources\runtime\cli\bin\dsh.cmd'
$Archive = (Resolve-Path './adenxie-dsh-vercel-0.3.1.tgz').Path
& $DshCli plugin --profile desktop add $Archive
Restart Desktop afterward. On another platform, use dsh plugin --profile web add /path/to/adenxie-dsh-vercel-0.3.1.tgz with your actual path and profile. To build from source instead:
Build the checkout before installing. Desktop and Web use separate profiles. For Desktop, initialize its profile once and fully quit the app, then run this in PowerShell:
git clone https://github.com/AdenXie/dsh-vercel.git
Set-Location dsh-vercel
pnpm install --frozen-lockfile
pnpm run build
$DshCli = Join-Path $env:LOCALAPPDATA 'Programs\DeepSeek Harness\resources\runtime\cli\bin\dsh.cmd'
$Checkout = (Get-Location).Path
& $DshCli plugin --profile desktop add "link:$Checkout"
These commands clone and build the source, then install the local checkout. Restart Desktop after installation. On other platforms, use the installed dsh CLI with dsh plugin --profile web add link:/path/to/dsh-vercel, replacing the path and profile for your setup. See the Chinese guide for authorization and current limitations. Exact supported versions are declared in package.json.
The package is scoped as @adenxie/dsh-vercel; DSH's Bundle row and settings panel use the unique ID adenxie-dsh-vercel. Its loopback-only API is under /api/adenxie/dsh-vercel/, and remote MCP tools use mcp__adenxie_vercel__* to avoid collisions with other Vercel plugins. Local helpers use the adenxie_vercel_* prefix. The browser module loader ID matches the scoped package name, as required by DSH's client module table.
I tested the scoped tarball in an isolated DSH profile. The real Host served the revisioned client bundle with a matching module ID; startup, disable, re-enable, and uninstall checks passed.
Fresh authorization requests openid offline_access so Vercel can issue a refresh token. Credentials use ~/.dsh/adenxie-dsh-vercel.json. On first load, if that file is absent, the plugin copies the legacy ~/.dsh/dsh-vercel.json to the new file and leaves the old file untouched; an existing new file always takes precedence. Existing credentials remain in place until reauthorization succeeds. If an older authorization has no refresh token, use the settings panel to authorize again. The loopback-only GET /api/adenxie/dsh-vercel/catalog endpoint returns the connected MCP tool catalogue, including mapped names, schemas, annotations, and metadata. Annotations are server hints and do not enforce access control. MCP structured results and any _meta are preserved in DSH's structured result fields; _meta is not rendered as model text.
The complete remote catalog stays registered with DSH, while each Agent initially sees only the catalog controls. Call adenxie_vercel_mcp_search_tools, then adenxie_vercel_mcp_load_tools with the returned names; the native Vercel tools become visible only in that Agent. Each Agent can load at most four tools with a combined schema budget of 32,000 characters. Use replace or adenxie_vercel_mcp_unload_tools to free capacity. A child Agent can load only tools its parent has loaded, and DSH host restrictions still apply. Disconnecting clears loaded schemas; search and load again after reconnecting.
Long tool results show the first 12,000 characters in model text and include a result handle in structured metadata when cached. Continue with adenxie_vercel_mcp_read_result, passing that resultId and the previous page's nextOffset; each page is limited to 12,000 characters. The cache holds at most eight results and one million characters per Agent, evicting the oldest entries when needed. Handles belong to the Agent that received the result and are cleared when the Agent is disposed or the MCP connection closes. Results too large to cache, agentless results, expired handles, and handles from another Agent cannot be continued; the tool response explains when you need to narrow the query.
On 2026-10-19, I checked the production plugin with the real DSH ToolRuntime and found 237 Vercel tools. The initial model schema was 2,986 characters for this plugin, compared with 488,972 characters for all 237 tools; the plugin schema footprint fell by 99.39%. Loading one tool exposed it only to Agent A, and the explicit-team list_projects call returned seven projects with verified ownership. The project-scope compatibility note describes that narrowly scoped read fallback. These schema measurements count serialized characters, not tokens, and cover this plugin's schemas rather than the whole application.
ToolRuntime, namespaced list_projects succeeded before and after an explicit connection restart, restoring all 237 tools. I also listed all 34 bundled skills and read their main files through the production loader. In Desktop chats, I verified model-driven reading of the deployments-cicd skill and the search → load → native project-query flow; selection in other tasks still depends on the model._meta survive for all 237 tools. The tested list_projects response itself had no _meta, so I checked response metadata preservation with a local MCP fixture. Automatic network-failure recovery remains fixture-only; my live check used an explicit restart.ctx.skills. It has no Vercel-specific pre-execution gate or embedded MCP App/widget UI; image and audio blocks are retained in structured results but do not receive native rendering.I develop and maintain this project as AdenXie. I based the initial MCP connection layer on zhengjy01/dsh-vercel-mcp and retain its MIT copyright and permission notice in LICENSE. I extended that foundation with scoped plugin identities, on-demand tool loading, bounded result paging, metadata preservation, and compatibility fixes.
The bundled Vercel skills are from vercel/vercel-plugin at commit 137e9f44a243a939c5eaea30e4dff96aee744f64 and are distributed under Apache-2.0. See LICENSES/ for their license and attribution.