Skip to content
/
Phase D Batch 27

Batch 27 — MCP Bridges & Skill Converters

Batch 27 — MCP Bridges & Skill Converters

Roster (10)

slug stars distribution cli_binary local_ui orchestration multi_model tier
gemini-cli-skill-foray 184 skill-pack no none hierarchical yes (claude+gemini) A
house-mcp-manager 8 cli-tool yes (house-mcp-manager) terminal-tui none no A
mcp-to-skill-converter 0 standalone-repo no (script) none none no A
claude-tools-mcp 13 mcp-server yes (go binary) none none no A
mcp2skill 51 cli-tool yes (mcp2skill) none none no A
claude-teams-mcp 271 mcp-server no tmux hierarchical yes (claude+opencode) A
pare-mcp 122 npm-package yes (npx init) none none no A
claude-context-local 231 mcp-server no none none no A
deebo 0 (fork) mcp-server no none parallel-fan-out yes (2 distinct models) A
code-mode-library 1,461 npm-package no none none no A

Intra-batch Patterns

All 10 frameworks address the same underlying problem — MCP tool definitions loading into context — but from radically different angles. House-mcp-manager, mcp-to-skill-converter, and mcp2skill all treat excessive MCP context load as a problem to solve (disable/convert); claude-flow seed treats it as a feature to maximize. The batch splits into three architectural clusters: (1) converters (mcp-to-skill-converter, mcp2skill — convert MCP to skills for lazy loading), (2) context managers (house-mcp-manager, claude-context-local, pare-mcp — control what's in context at runtime), and (3) new protocols (claude-tools-mcp, claude-teams-mcp, deebo, code-mode-library — create new ways to structure agent-tool interaction entirely).

The gemini-cli-skill-foray stands alone as the only framework in this batch concerned with cross-model routing rather than MCP/skill/context management. Multi-model support is present in three frameworks (gemini-cli-skill-foray, claude-teams-mcp, deebo) but each implements it differently: gemini-cli delegates to a second LLM via subprocess, claude-teams mixes claude/opencode backends, and deebo uses different models per agent role.

Most Interesting Finds

  1. code-mode-library (1,461 stars): The most radical rethinking of tool calling in the entire corpus — replacing all MCP tools with a single TypeScript code execution interface, citing Apple/Cloudflare/Anthropic research, with independent benchmark validation showing 88% efficiency gains for complex workflows. This is not an incremental improvement; it's a paradigm shift.

  2. deebo: The only framework explicitly designed for hypothesis-driven parallel debugging, using git-branch isolation (not worktrees, not containers) per agent. The architecture where session logs are designed to be consumed by an AI agent (not humans) for status checking is a novel pattern. The framework is dormant but spawned a commercial successor (bojack.ai).

Items Written as Tier C

None. All 10 frameworks had sufficient public material for full 11-file reports.

Cross-References Discovered

  • mcp-to-skill-converter (GBSOSS/Python) and mcp2skill (fenwei-dev/Go) implement the same concept at different quality levels. Both cite the "playwright-skill progressive disclosure pattern" as inspiration. mcp2skill is the production successor.
  • claude-teams-mcp explicitly reverse-engineers Claude Code's native agent teams protocol — making it a downstream artifact of Claude Code's internal design rather than an independent invention.
  • deebo original repo (snagasuri/deebo-prototype) is no longer accessible at the assigned URL; the MjrTom fork was used for analysis. The note about the URL in the batch assignment should be updated if this research is referenced.
  • house-mcp-manager is part of a three-tool "House suite" (house-code, house-agents, house-mcp-manager). The other two tools were not in this batch.
  • code-mode-library is part of the broader UTCP (Universal Tool Calling Protocol) ecosystem, which proposes an alternative to MCP. The @utcp/mcp plugin within it bridges back to MCP, making it both a UTCP tool and an MCP-compatible tool.