Rovodev CLI MCP
Exposes Rovodev CLI as MCP tools for interacting with Rovodev agent, including session management, streaming chunk-caching, and tool-based CLI operations.
README
Rovodev CLI MCP
An MCP server wrapper that exposes Rovodev CLI as MCP tools. Modeled after gemini-mcp-tool.
Install
npm install -g rovodev-cli-mcp
Or use with npx:
npx -y rovodev-cli-mcp
Configure in Claude Desktop
{
"mcpServers": {
"rovodev": {
"command": "npx",
"args": ["-y", "rovodev-cli-mcp"]
}
}
}
Environment variables to customize underlying CLI:
ROVODEV_CLI_PATHorROVODEV_CMD(default:acli)ROVODEV_SUBCOMMAND(default:"rovodev run")ROVODEV_CONFIG_FLAG(default:--config-file)ROVODEV_SHADOW_FLAG(default:--shadow)ROVODEV_VERBOSE_FLAG(default:--verbose)ROVODEV_RESTORE_FLAG(default:--restore)ROVODEV_YOLO_FLAG(default:--yolo)ROVODEV_HELP_FLAG(default:--help)
Chunking configuration (for large responses):
MCP_CHUNK_SIZE(preferred)ROVODEV_CHUNK_SIZE(legacy) The first one found is used; default is 20000 characters if none are set.
Global setup for Claude Code (user scope)
Install the MCP server for your user so it’s available across projects in Claude Code.
-
Add with npx (no global npm install required):
claude mcp add -s user rovodev -- npx -y rovodev-cli-mcp -
Optional: use a globally installed binary instead of npx
npm i -g rovodev-cli-mcp claude mcp add -s user rovodev rovodev-cli-mcp -
Optional: set environment overrides when adding
claude mcp add -s user rovodev -- \ npx -y rovodev-cli-mcp \ -e ROVODEV_CLI_PATH=acli \ -e ROVODEV_SUBCOMMAND="rovodev run" \ -e ROVODEV_CONFIG_FLAG=--config-file \ -e ROVODEV_SHADOW_FLAG=--shadow \ -e ROVODEV_VERBOSE_FLAG=--verbose \ -e ROVODEV_RESTORE_FLAG=--restore \ -e ROVODEV_YOLO_FLAG=--yolo \ -e ROVODEV_HELP_FLAG=--help -
Verify it’s connected:
claude mcp list -
Remove from user scope (to update or uninstall):
claude mcp remove -s user rovodev
Global setup for Amazon Q CLI
Install the MCP server for Amazon Q CLI to use Rovodev tools in Q chat sessions.
-
Add with npx (no global npm install required):
q mcp add --name rovodev --command npx --args "-y,rovodev-cli-mcp" -
Optional: set environment variables when adding
q mcp add --name rovodev --command npx --args "-y,rovodev-cli-mcp" \ --env "ROVODEV_CLI_PATH=acli" \ --env "ROVODEV_SUBCOMMAND=rovodev run" -
Verify it's installed:
q mcp list -
Check server status:
q mcp status --name rovodev -
Remove (to update or uninstall):
q mcp remove --name rovodev
Using tools in Q CLI: Tools are available with the rovodev___ prefix:
rovodev___ask-rovodev- Send messages to Rovodev agent (enhanced with session support)rovodev___tap-rovodev- Alias for ask-rovodevrovodev___session_manager- Manage isolated sessions for context persistencerovodev___health_check- Check server health and CLI availabilityrovodev___diagnostics- Get comprehensive system diagnosticsrovodev___next-chunk- Fetch next chunk for large responsesrovodev___fetch-chunk- Fetch specific chunk by indexrovodev___Help- Show Rovodev CLI helprovodev___Ping- Test connectivity
Local development with Claude Code
If you are iterating on this repo and want Claude Code to use your local build:
- Install deps and build
npm i
npm run build
- Point Claude Code at your local
dist/index.js
claude mcp remove -s user rovodev || true
claude mcp add -s user rovodev -- node "$(pwd)/dist/index.js"
- Verify connection
claude mcp list
- Optional: quick smoke test (no Claude required)
node scripts/smoke.mjs
Notes:
- This server communicates over stdio. Do not print to stdout; only stderr is safe for logs.
- The server already routes logs to stderr; if you add logs, follow the same pattern.
- You can still register via
npx -y rovodev-cli-mcp, but during development usingnode dist/index.jsavoids npm network hiccups.
Troubleshooting
-
"Failed to connect" in
claude mcp list- Ensure your Node is >= 18 (
node -v). - Ensure no output is written to stdout (stdout is reserved for the MCP protocol). This repo's logger writes to stderr only.
- Try registering the local path:
claude mcp add -s user rovodev -- node "$(pwd)/dist/index.js". - If using WSL, register from the Linux side; Claude Code on Windows can discover WSL user-scoped config.
- Ensure your Node is >= 18 (
-
Authentication errors (API key issues)
- If you see "Authentication Error" messages, your rovodev CLI API key may be expired or invalid
- Check your rovodev CLI configuration and update your API key
- The server now detects authentication issues and provides clear error messages instead of hanging
-
Underlying CLI not found
- The default command is
acli. Install it or override:
claude mcp add -s user rovodev -- \ node "$(pwd)/dist/index.js" \ -e ROVODEV_CLI_PATH=/path/to/your/cli - The default command is
-
Large responses
- Use
ask-rovodevfirst. If the response is chunked, callnext-chunkrepeatedly with the providedcacheKey. You can also fetch a specific page withfetch-chunk. - Tune chunk size with env vars (in order of precedence):
MCP_CHUNK_SIZE, then legacyROVODEV_CHUNK_SIZE.
- Use
Tools and streaming
This server supports streaming chunk-caching to reduce memory usage for large outputs. The ask-rovodev tool streams the CLI output directly into a chunk cache. If the output exceeds the page chunk size, it returns the first chunk and a cacheKey so you can fetch subsequent chunks with next-chunk or fetch-chunk.
Tools
Enhanced Core Tools
ask-rovodev:{ message?: string, prompt?: string, configFile?: string, shadow?: boolean, verbose?: boolean, restore?: boolean, yolo?: boolean, sessionId?: string, retries?: number, backoffMs?: number, args?: string[], pagechunksize?: number }- New: Session isolation with
sessionId, retry logic withretries/backoffMs - Default:
yolomode is enabled for non-interactive MCP usage; setyolo: falseto disable if you need interactive behavior. - Tip: If your message starts with dashes (e.g.,
--example), the underlying CLI may interpret it as a flag. The server inserts--before such messages to prevent flag parsing.
- New: Session isolation with
tap-rovodev: alias ofask-rovodev
Session Management
session_manager:{ action: "create" | "destroy" | "list" | "get" | "cleanup", sessionId?: string, timeoutMs?: number }- Manage isolated sessions for maintaining context across multiple commands
Health & Diagnostics
health_check:{ detailed?: boolean }- Check server health including CLI availability, session manager status, and environment configuration
diagnostics:{ includePerformance?: boolean, includeHealth?: boolean, includeSessions?: boolean, includeSystem?: boolean, includeRecentExecutions?: boolean, recentLimit?: number }- Get comprehensive system diagnostics including performance metrics and resource usage
Utility Tools
next-chunk:{ cacheKey: string }Fetch the next chunk sequentially from a cached large responsefetch-chunk:{ cacheKey: string, chunkIndex: number }(optional) Fetch a specific chunk by indexHelp: show help from underlying CLIPing: connectivity test
Server behavior and tuning via env vars:
- Logging level:
MCP_LOG_LEVEL(debug | info | warn | error | silent) - Exec timeout:
MCP_EXEC_TIMEOUT_MS(kill underlying CLI after N ms; includes partial stdout tail in error) - Max stdout buffer:
MCP_MAX_STDOUT_SIZE(maximum output buffer size before termination) - Working directory:
MCP_CWD(set process cwd for the underlying CLI) - Chunk cache TTL:
MCP_CHUNK_TTL_MS(default 20 minutes) - Chunk cache max entries:
MCP_CHUNK_MAX_ENTRIES(default 500)
Handling large responses (simple sequential flow)
- If the output is too large, the first
ask-rovodevcall returns page 1 plus acacheKey:- Response header example:
cacheKey: <key>,chunk: 1/N
- Response header example:
- To continue, call
next-chunkwith the samecacheKeyto get chunk 2, then 3, etc. until it reports no further chunks.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。