devtap
Bridge build/dev process output to AI coding sessions via MCP — supports Claude Code, Codex, OpenCode, Gemini CLI, and aider
claude mcp add --transport stdio killme2008-devtap docker run -i killme2008/devtap
How to use
devtap is a tool that bridges build and development command outputs to AI coding sessions via MCP (Model Context Protocol). It captures stdout and stderr from your build and dev commands and fans that output to multiple AI adapters (Claude Code, Codex, OpenCode, Gemini, etc.), allowing each agent to drain its own copy for parallel analysis. The server side (MCP integration) is driven by the devtap tool, which can operate in local mode with file-based queues or connect to a shared store like GreptimeDB for cross-machine setups. With devtap installed, you can configure adapters for different AI tools, pipe build logs into those adapters, and even query on-demand errors via the built-in get_build_errors skill. In practice, you run a development command in one terminal (for example, cargo check or npm run build), and devtap takes the output and distributes it to all configured AI tools so they can propose fixes, summarize failures, or proactively suggest improvements.
How to install
Prerequisites:
- Go tooling (required to install devtap, as per the Quick Start in the README)
- GitHub access to fetch the module
Installation steps:
- Install devtap (Go):
go install github.com/killme2008/devtap/cmd/devtap@latest
- Optionally install via Homebrew (macOS):
brew install killme2008/tap/devtap
- Verify installation:
devtap version
Note: The MCP server feature integrates through the devtap tool itself; you will configure MCP adapters and routes using devtap commands (e.g., devtap install, devtap mcp-serve) after the binary is available.
Additional notes
Tips and considerations:
- The storage backends include a local file queue by default and an optional GreptimeDB setup for persistent history and richer querying. If you enable cross-machine usage, ensure your store is reachable by all clients.
- Each adapter gets its own queue (pending.jsonl) to avoid cross-talk between tools.
- The MCP server, as described in the README, supports multiple tools and can funnel the same output to several agents in parallel.
- If you want to test MCP without connecting adapters, you can use devtap drain to verify the captured output formatting and routing.
- When using cross-machine mode, you can enable different sessions or stores; DevTap will merge sources if the same backend+session are detected, otherwise it will drain and prefix sources accordingly.
- Environmental variables and config options can be adjusted in the .mcp.json/.codex/.claude-like config files as described by the specific adapters.
Related MCP Servers
grepai
Semantic Search & Call Graphs for AI Agents (100% Local)
agent-deck
Terminal session manager for AI coding agents. One TUI for Claude, Gemini, OpenCode, Codex, and more.
claude-code-open
Open source AI coding platform with Web IDE, multi-agent system, 37+ tools, MCP protocol. MIT licensed.
floop
Spreading activation memory for AI coding agents - corrections in, context-aware behaviors out.
mcp-jest
Automated testing for Model Context Protocol servers. Ship MCP Servers with confidence.
mcp-tidy
CLI tool to visualize and manage MCP server configurations in Claude Code. List servers, analyze usage statistics, and clean up unused servers