refactor: consolidate MCP tools from 15 to 10 using manage_* pattern#89
refactor: consolidate MCP tools from 15 to 10 using manage_* pattern#89masnwilliams merged 8 commits intomainfrom
Conversation
…em, proxies, extensions, and process execution Brings the MCP server up to parity with the latest Kernel API. Updates @onkernel/sdk from ^0.30.0 to ^0.35.0. New tools: - Browser Pools: create, list, get, acquire, release, delete - Computer Controls: click, type_text, press_key, scroll, move_mouse, get_mouse_position - Proxies: create, list, delete - Extensions: list, delete - Filesystem: read_file, write_file, list_files - Process: exec_command New resource: browser_pools:// Co-authored-by: Cursor <cursoragent@cursor.com>
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
Applies MCP best practices from Block, Anthropic, and ecosystem research. LLM accuracy degrades after 20-25 tools; we now have 10. 6 manage_* tools (one per feature domain): - manage_browsers: create/list/get/delete - manage_profiles: setup/list/delete - manage_browser_pools: create/list/get/delete/flush/acquire/release - manage_proxies: create/list/delete - manage_extensions: list/delete - manage_apps: list_apps/invoke/get_deployment/list_deployments/get_invocation 4 standalone tools: - computer_action: click/type/press_key/scroll/move/get_position/screenshot - execute_playwright_code (unchanged) - exec_command (new, replaces read_file/write_file/list_files) - search_docs (unchanged) Co-authored-by: Cursor <cursoragent@cursor.com>
masnwilliams
changed the title
The executable goes in `command`, arguments go in `args` (e.g., command: "cat", args: ["/etc/resolv.conf"]). Co-authored-by: Cursor <cursoragent@cursor.com>
…e mapping Agents receive scaled-down screenshot images but need to use the actual viewport coordinate space (default 1920x1080) for click/scroll/move. The screenshot action now fetches browser info in parallel and returns the viewport dimensions alongside the image. Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
![cursor[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
Co-authored-by: Cursor <cursoragent@cursor.com>
| const blob = await screenshotResponse.blob(); | ||
| const buffer = Buffer.from(await blob.arrayBuffer()); | ||
| const viewport = browserInfo.viewport; | ||
| return { |
There was a problem hiding this comment.
computer_action screenshot can crash on missing browser
Medium Severity
In computer_action action "screenshot", browserInfo from client.browsers.retrieve is assumed non-null and browserInfo.viewport is accessed unguarded, so a null/undefined retrieve result can throw after the screenshot succeeds.
Restore the detailed SSH port forwarding explanations, activity timeout warning, and install links that were dropped during the consolidation. Re-add partial region argument validation for computer_action screenshot so it errors instead of silently falling back to full screenshot. Co-authored-by: Cursor <cursoragent@cursor.com>
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes and found 8 potential issues.
Bugbot Autofix is OFF. To automatically fix reported issues with Cloud Agents, enable Autofix in the Cursor dashboard.
| /// <reference types="next" /> | ||
| /// <reference types="next/image-types/global" /> | ||
| import "./.next/types/routes.d.ts"; | ||
| import "./.next/dev/types/routes.d.ts"; |
There was a problem hiding this comment.
Auto-generated file incorrectly modified with dev-only path
Medium Severity
The next-env.d.ts file references .next/dev/types/routes.d.ts, which is generated only during next dev and filtered out during builds. Next.js 16 generates both .next/dev/types (dev mode) and .next/types (build mode), with production builds expecting the latter. This file explicitly states it should not be edited and appears to be an accidentally committed auto-generated file from a dev environment, potentially causing TypeScript errors during production builds.
Co-authored-by: Cursor <cursoragent@cursor.com>
1 participant
Summary
Rewrites the MCP server's tool set based on research from Block (60+ MCP servers), Phil Schmid, Docker, and Anthropic on MCP best practices:
actionparameterThis PR consolidates 15 existing tools + adds new capabilities = 10 total tools using a consistent
manage_*pattern, plus upgrades@onkernel/sdkto^0.35.0.The 10 tools
manage_ tools (one per feature domain)*
manage_browsersmanage_profilesmanage_browser_poolsmanage_proxiesmanage_extensionsmanage_appsStandalone tools
computer_actionexecute_playwright_codeexec_commandsearch_docsNew capabilities added
New resource
browser_pools://-- list/get browser poolsTest plan
tsc --noEmit)Note
Medium Risk
Large surface-area API change: tool names/signatures and workflows change substantially and could break existing MCP client integrations. New VM command execution and computer-control capabilities increase operational impact if misused, though still gated by auth.
Overview
Refactors the MCP tool API from many endpoint-style tools into 10 total tools centered on
manage_*tools with anactionparameter, replacing the prior browser/profile/app/deployment/invocation tool set with consolidatedmanage_browsers,manage_profiles,manage_apps, plus newmanage_browser_pools,manage_proxies, andmanage_extensions.Adds new capabilities: a
computer_actiontool for mouse/keyboard controls and screenshots (supersedingtake_screenshot), anexec_commandtool to run shell commands inside the browser VM with decoded stdout/stderr, and a newbrowser_pools://resource for listing/retrieving pools. Updatessearch_docserror handling/formatting, refreshes docs/examples/troubleshooting language to the new tool names, and bumps@onkernel/sdkto^0.35.0(plusnext-env.d.tsroute types path update).Written by Cursor Bugbot for commit f7b13f7. This will update automatically on new commits. Configure here.