Detailed Analysis
A developer building in the Claude Code ecosystem has released Soubi, an open-source compiler that translates Claude Code plugins—skills, hooks, subagents, commands—into native formats for 22 different AI coding harnesses, including Cursor, Codex, OpenCode, Gemini CLI, GitHub Copilot, and Windsurf. The tool addresses a specific pain point that has emerged as Claude Code's plugin ecosystem has matured: developers who build extensions for Anthropic's coding agent find that format is essentially locked to that one environment, forcing manual, harness-specific rewrites to reach users on competing tools. Soubi's pitch, explicitly framed as "Next.js for agent plugins," lets a developer author a plugin once in a standard directory structure and then run a build command that outputs the correct native artifact for each target—a real plugin manifest with MCP support for Claude Code, Agent Skills plus TOML-based agents for Codex, a marketplace manifest with rules and MCP config for Cursor, and so on.
The more substantive engineering problem the project tackles is semantic mismatch rather than simple format translation. Claude Code exposes roughly 30 distinct hook events, while most other harnesses support only a handful, a subset of handler types, or none at all. Rather than silently dropping incompatible features—which would ship broken or hollow plugins to end users—Soubi requires each capability to declare an explicit fallback behavior: degrade gracefully to a best-effort equivalent, drop with a warning, or hard-exclude that harness from the build entirely when a feature is load-bearing to the plugin's core function. The example given is subagent delegation: if a harness has no concept of subagents, there is no meaningful compilation target, so the build fails loudly rather than producing a plugin that appears to work but silently omits its central mechanism. This design choice reflects a broader lesson many cross-platform tooling projects have learned the hard way—that graceful degradation without transparency erodes trust faster than an explicit build failure.
This development is notable as a symptom of how quickly Claude Code's plugin system has become a de facto standard that other tools are implicitly measured against. Anthropic's decision to ship a rich extensibility model—skills, hooks, subagents, commands, and MCP integration—inside Claude Code created an ecosystem effect: plugin authors invested effort in Anthropic's format first, and now a third party sees enough value in that ecosystem to build tooling that exports it elsewhere. That dynamic mirrors what happened earlier in the web framework and infrastructure-as-code spaces, where a dominant platform's configuration format becomes the schema everyone else needs an adapter for. It's also a signal that the AI coding agent market is fragmenting across a genuinely large number of harnesses (Soubi targets 22), each with its own plugin conventions, which creates real friction for tool authors trying to reach the broadest audience without maintaining parallel codebases.
More broadly, this fits into a pattern seen throughout 2025 and into 2026 of the AI coding tool ecosystem maturing past the "single vendor, single format" phase and into an interoperability phase, similar to how the Model Context Protocol (MCP) itself emerged from Anthropic and was subsequently adopted across competing platforms as a shared standard for tool and data connections. Community-built compilers, adapters, and shims like Soubi suggest that developers are not waiting for a formal cross-vendor standard for plugins the way MCP established one for tool-calling—instead, grassroots tooling is emerging to paper over fragmentation in real time. For Anthropic, this is a mixed signal: it validates Claude Code's plugin architecture as sufficiently rich and well-designed to be worth porting elsewhere, while also underscoring that plugin authors want reach beyond the Claude Code user base, which could eventually pressure Anthropic or the broader ecosystem toward a more formal, shared plugin specification akin to what MCP achieved for tool integration.
Read original article →