Google stitch-skills: Agent Skills for the Stitch MCP Server
WHY IT MATTERS
Google Labs published stitch-skills, a library of agent skills that follow the Agent Skills open standard and target the Stitch MCP server, compatible with Antigravity, Gemini CLI, Claude Code, and Cursor.
What Happened
Google Labs published stitch-skills, a repository of agent skills built to the Agent Skills open standard and targeting the Stitch MCP server. The library is documented as compatible with Antigravity, Gemini CLI, Claude Code, and Cursor, and is distributed via GitHub under Google Labs' namespace. The release packages reusable agent capabilities as portable artifacts rather than runtime-specific prompts or plugins.
Why It Matters
Skill portability across agent runtimes has been an unresolved integration cost: teams building on Antigravity, Gemini CLI, Claude Code, or Cursor have historically maintained parallel prompt libraries, tool bindings, and invocation conventions per platform. stitch-skills anchors the Stitch MCP server as a reference target for a vendor-neutral skill format, which converts skill authoring from a per-runtime maintenance burden into a single artifact that can be validated once and deployed across four distinct harnesses. For operators running heterogeneous agent fleets, this reduces drift between environments and simplifies audit surface — one skill definition, one behavior specification, multiple execution contexts. Google publishing under an open standard rather than a proprietary format also signals that skills, like MCP servers before them, are being treated as infrastructure primitives with expected cross-vendor circulation. The practical beneficiary is any team that has delayed standardizing on a skill library because of runtime lock-in concerns.
Technical Details
The library conforms to the Agent Skills open standard, meaning skills are packaged as structured artifacts (metadata plus invocation logic) rather than freeform prompts, and are consumed by the host runtime's skill loader. Targeting the Stitch MCP server means these skills exercise Stitch capabilities through MCP tool calls, so the skill definitions depend on MCP server availability and versioning rather than on any single agent's native tool interface. Compatibility spans four runtimes with materially different execution models — Antigravity's agent orchestration, Gemini CLI's terminal-driven flow, Claude Code's file-and-shell-centric loop, and Cursor's IDE integration — which implies the skill format abstracts invocation semantics sufficiently to survive those differences. The standard constrains what a skill can express: skills describe capability and tool usage, not runtime-specific control flow, so complex behaviors requiring host-specific hooks will still need per-platform adaptation. Distribution via GitHub means versioning follows repository releases, with no managed registry or signing layer described in the initial publication.
Operational Impact
Builders can now author a skill once and validate it against four runtimes without rewriting invocation logic, cutting the marginal cost of supporting a second or third agent platform. Operators gain a consistent behavioral contract across environments: the same skill against the same Stitch MCP server should produce comparable tool-call sequences regardless of host, which simplifies regression testing and incident triage. Prompt libraries maintained per-runtime become candidates for consolidation, and internal skill repositories can standardize on the Agent Skills format as their interchange layer. The dependency to manage shifts from runtime-specific prompt drift to MCP server versioning — upgrading Stitch or changing its tool surface now has fleet-wide skill implications. Teams without an MCP integration layer will find less value here; the release presumes MCP as the tool substrate.
SOURCE
GitHub
SHARE
MORE FROM STUFFINSIDER
Prefect Releases FastMCP: Python Framework for MCP Servers
Nov 15DEVELOPER TOOLSTencent LoopForge: Self-Hosted Coding-Agent Delivery Workflows
Nov 15DEVELOPER TOOLSHugging Face and OpenRouter Partner to Simplify Open-Weight Model Inference
Nov 16DEVELOPER TOOLSApple Tests Daily Limits for Complex Siri AI Interactions
Nov 16