routinekit

dsh plugin --profile web add routinekit

RoutineKit

Your agent did it once. Make it a tool you can run again.

RoutineKit turns successful, explicitly selected tool calls into reviewed, parameterized, executable routines. Change the inputs, check the live tools, approve the run, and replay without asking a model to reconstruct each step.

A scripted capture/replay tutorial through the official filesystem MCP server: a new file path, new content, and a read-back success check. Watch the recording.

Start in DeepSeek Harness (DSH), or use the standalone workbench and MCP server with native WebMCP pages. Export a routine with a companion SKILL.md for another agent to discover and run.

v0.2 preview. Connect an explicitly configured local MCP server, capture its calls, Save as Tool, and reuse typed inputs. The workbench also runs as an MCP App in compatible hosts. This is a deterministic runner for reviewed linear workflows, not a general computer-use agent or OS sandbox.

Try it locally

Requires Node 22.19+. Native WebMCP additionally needs a current Chrome/Edge or Playwright Chromium. Tested locally with Chromium 151 and Chrome 152; browser APIs are experimental and may change. Local MCP file workflows do not require a WebMCP-capable browser.

npx routinekit doctor
npx routinekit demo

The demo opens a local workbench beside a real browser playground. A disclosed scripted tutorial, not an LLM, has made two real WebMCP calls: find a sample project and open its returned id.

1. Click Preview capture. Notice the input binding and the reference to the first step's returned id.

2. Click Add success check, choose step_2 /opened, and enter true. Check Save as a named tool, then Review & save and approve the exact routine. The JSON editor is optional.

3. Enter timer (or drawing) as the new category, then Review & run.

4. Approve the run and each WebMCP call. The selected project changes: the second step uses the new result id, not the recorded one.

The demo's localhost origin exists only while that process runs. Demo routines deliberately do not silently rebind to a different origin after a restart. For persistent use, connect your own WebMCP app at a stable origin.

No account, model API key, or GPU is needed for the tutorial. If no supported browser is installed:

npx playwright-core@1.62.1 install chromium

Try a real filesystem MCP workflow

In a new empty demo directory, explicitly install the reference server and RoutineKit:

npm init -y
npm install routinekit @modelcontextprotocol/server-filesystem@2026.8.31
npx routinekit demo-files ./node_modules/@modelcontextprotocol/server-filesystem/dist/index.js

This disclosed scripted tutorial starts the installed official filesystem MCP server with a new temporary workspace, writes first-note.txt, and reads it back. It does not use your existing files. In the workbench:

1. Add the check step_2 /content compared with Input: content.

2. Select Save as a named tool, review the bindings, and approve saving.

3. Use the new file path printed in the terminal, enter different content, and run with approval. The result must equal the new content.

4. The same routine is available as routine_saved_write_and_check with direct file and content inputs in its owning MCP session. Choose Export skill ZIP to keep it.

Exiting removes the demo's temporary files and routines. An exported routine needs a freshly approved matching files server before it can run elsewhere; permissions and file access are not transferred. The capture is scripted, not evidence of autonomous model success.

In DeepSeek Harness

dsh plugin --profile web add routinekit

Use DSH 0.1.2-rc.1 for this preview. The plugin registers tools and adds a RoutineKit button to the web sidebar. It does not modify the DSH core or replace your existing tool permissions.

Open the panel and choose Initialize RoutineKit in DSH, or ask:

Use routine_tools to list the tools you can record. I want to repeat [my task] with [these changing inputs]. Before doing it, use routine_record with only the needed tools and exact example argument values. Complete the task, show me routine_preview, and propose checks that prove its result. Save only after I approve.

When replaying a DSH-native routine, the panel's Run in DSH button queues a routine_run request through the current conversation. That launch can use the host model. The replay engine itself makes zero model calls and dispatches each step through ctx.tools.execute, retaining the same agent scope, cancellation, approval policies, and guards.

Recording is task-local. Saved DSH routines are kept under DSH_HOME/routinekit (partitioned by the available workspace/task identity); standalone storage uses ROUTINEKIT_HOME or ~/.routinekit. Explicitly exposed saved tools are registered in the owning agent scope, never globally for every workspace. Saving a routine does not grant permission to run it.

The web panel is loopback-only in this preview. Remote/LAN access is not supported. If DSH's package-age policy rejects a just-published package, wait for that policy window; do not lower it to install RoutineKit.

In an MCP client

Add this stdio server using your client's normal MCP configuration:

{
  "mcpServers": {
    "routinekit": {
      "command": "npx",
      "args": ["-y", "routinekit", "mcp"]
    }
  }
}

Human form elicitation is required for actions. A client without it can list/inspect routines but cannot auto-approve execution. Use npx routinekit open for the standalone workbench instead. Do not infer full clie

…(truncated)