Back to Directory/File Systems

io.github.Lmeng486/nano-mcp-termux

Zero-dependency single-file MCP server for Termux/Android, Raspberry Pi and Linux

File SystemsPythonv1.1.1

nano-mcp-termux

Zero-dependency, single-file MCP server for AI agents, natively running on Termux (Android), Raspberry Pi, and Linux.

GitHub Repository License: MIT MCP Version Docker Ready Version Tests CI Zero Dependencies

🚀 Quick Start

Local (Termux / Linux)

cd /path/to/nano-mcp-termux
python3 nano_mcp_termux.py

Docker

docker build -t nano-mcp-termux .
docker run -i --rm nano-mcp-termux

Health check (one-shot, no MCP handshake)

python3 nano_mcp_termux.py --health

🔌 Connect to your MCP client

This server does not run standalone — it's a sidecar that your AI client connects to via the MCP protocol.
Below are the three most common ways to mount it.

Option 1 — Zero‑install (uvx), best for desktop AI clients

uvx nano-mcp-termux

Then add this JSON to your AI client's mcpServers config (see client‑specific notes below):

{
  "mcpServers": {
    "nano-mcp-termux": {
      "command": "uvx",
      "args": ["nano-mcp-termux"]
    }
  }
}

Option 2 — Already pip installed, or you prefer a direct Python launch

pip install nano-mcp-termux

Config snippet for your AI client:

{
  "mcpServers": {
    "nano-mcp-termux": {
      "command": "nano-mcp-termux"
    }
  }
}

Option 3 — Run the single file directly (e.g. inside Termux on Android)

# Clone or download the repo, then:
cd /path/to/nano-mcp-termux
python3 nano_mcp_termux.py

Config snippet for your AI client:

{
  "mcpServers": {
    "nano-mcp-termux": {
      "command": "python3",
      "args": ["/full/path/to/nano_mcp_termux.py"]
    }
  }
}

How to use these snippets: client‑specific config files

AI clientConfig file location (examples)Note
Claude DesktopmacOS: ~/Library/Application Support/Claude/claude_desktop_config.json
Windows: %APPDATA%\Claude\claude_desktop_config.json
∆ mcpServers key at top level
Cursor~/.cursor/mcp.json∆ mcpServers key at top level
Cline$HOME/.cliner/mcp.json (or inside project folder as cline_mcp_settings.json)∆ mcpServers key at top level
Cherry StudioGUI → MCP section (typically auto‑creates ~/.cherrystudio/mcp/*.json)JSON schema as above
Operit~/.operit/mcp_config.json (or per‑workspace config in the workspace folder)∆ mcpServers key at top level

Tip: after editing the config file, restart your AI client so the new server is detected. If the client supports hot‑reload, you may not need a restart.

Troubleshooting

SymptomLikely causeFix
Client shows "server not found"command path is wrong or not in PATHUse absolute path (/usr/local/bin/nano-mcp-termux for pip installs, $HOME/.local/bin/nano-mcp-termux for pip --user)
Error: spawn uvx ENOENTuvx not in PATH (not installed)Use a custom shell like "shell": "uw uvx" or the direct Python launch (Option 2)
"Connection lost" / process dies immediatelyNo stdin from clientThis server is stdio‑bound; don't run it directly like nano-mcp-termux — always launch it through your client's MCP config

🛠️ Features

  • ✅ Zero-dependency: No external packages required
  • ✅ Single-file: All logic in one Python script
  • ✅ Cross-platform: Works in Termux (Android), on Raspberry Pi, Linux, and inside Docker
  • ✅ Container-ready: Dockerfile at repo root for Glama introspection
  • ✅ Standard stdio framing: Newline-delimited JSON-RPC (official MCP transport)
  • ✅ Version negotiation: Echoes client protocolVersion (2024-11-05 / 2025-06-18 / 2025-11-25), falls back to highest supported
  • ✅ Graceful degradation: device_status never fails; unavailable inputs are marked, not crashed
  • ✅ [lifecycle] No orphaned processes: every command runs in its own process group; timeouts kill the whole tree (SIGTERM → SIGKILL), never just the shell
  • ✅ [lifecycle] Self-terminating: parent-process watchdog + stdin-EOF detection + optional idle timeout, so a dead client never leaves a zombie server behind
  • ✅ [lifecycle] Phantom Process Killer aware: reports Android 12+ child-count risk with concrete mitigations
  • ✅ [paths] Android/Termux path adapter: resolves /sdcard ↔ /storage/emulated/0 ↔ $HOME/storage/shared, detects noexec mounts, case-insensitive filesystems, and shell-quoting hazards

🧩 Tools Overview

ToolDescription
device_statusAggregate device info: battery / storage / arch / memory (degraded fields marked unavailable)
server_healthRuntime health: pid/ppid/env, uptime, request counters, live children, reaping policy, RSS, Phantom Process Killer risk
resolve_pathResolve an Android/Termux path: canonical form, mount point, noexec/read-only flags, case sensitivity, equivalent spellings, quoting warnings
path_adapterEnvironment profile: runtime class, $HOME/$PREFIX/$TMPDIR, Android storage roots with mount summaries, noexec mount list

🔗 Protocol Surface

MethodBehavior
initializeEchoes supported client version (2024-11-05 / 2025-06-18 / 2025-11-25), else responds with highest supported per MCP spec; declares {"tools": {}}
notifications/initializedSilent (no response)
pingReturns {}
tools/listReturns tool catalog
tools/callReturns {"content": [{"type": "text", ...}]}
Unknown method / tool-32601 / -32602 JSON-RPC errors
Malformed JSON-32700 parse error

🔄 Why 1.1.0: the orphan-process bug

Version 1.0.0 ran commands with subprocess.run(..., timeout=...). On timeout, Python terminates only the direct child — the sh -c wrapper. Anything the shell had already spawned is reparented to init and survives forever. On Android that is the exact leak class that trips the Phantom Process Killer and can get the whole app group killed.

1.1.0 fixes this at the root: commands are spawned with start_new_session=True and torn down with os.killpg(pgid, SIGTERM) → SIGKILL, so the entire process tree dies together.

The repository ships a reproducible A/B regression test for it:

python3 test_lifecycle_orphan.py
# current  nano_mcp_termux.py           leaked_grandchildren=0
# legacy   nano_mcp_termux_v10_backup.py leaked_grandchildren=1   <- the bug
# PASS: current server leaves 0 orphaned grandchildren after timeout

✅ Tests

python3 test_client.py            # 32 protocol / tool / lifecycle assertions
python3 test_lifecycle_orphan.py  # orphan-leak A/B regression (1.0.0 vs 1.1.0)

⚙️ Environment variables

VariableDefaultPurpose
NANO_IDLE_TIMEOUT0 (off)Seconds of inactivity before the server exits on its own
NANO_NO_WATCHDOGunsetSet to 1 to disable the parent-process watchdog thread

🔗 Links


This server is designed to pass Glama introspection checks and serve as a lightweight MCP implementation for mobile and edge devices.

Installation

Source-derived launch command. Check the maintainer’s required arguments and credentials before running:

bash
uvx nano-mcp-termux

Set up in your AI client

Merge this template into ~/Library/Application Support/Claude/claude_desktop_config.json. Keep existing servers. Add any arguments, credentials, and permissions required by the maintainer; this template has not been install-tested.

json
{
  "mcpServers": {
    "io-github-lmeng486-nano-mcp-termux": {
      "command": "uvx",
      "args": [
        "nano-mcp-termux"
      ]
    }
  }
}

Restart Claude Desktop completely for changes to take effect. Confirm the server appears connected in the client’s tool list, then try a read-only example from its documentation.

Claude Desktop setup reference

Package

nano-mcp-termuxpypi

Compatible MCP Clients

io.github.Lmeng486/nano-mcp-termux works with any MCP-compatible client. Copy the config snippet from the Configuration section above and add it to the file shown for your client, then restart the application.

  • Claude Desktop~/Library/Application Support/Claude/claude_desktop_config.jsonRestart Claude Desktop completely for changes to take effect.
  • Cursor~/.cursor/mcp.jsonRestart Cursor for changes to take effect.
  • VS Code.vscode/mcp.jsonReload VS Code window for changes to take effect.
  • Windsurf~/.codeium/windsurf/mcp_config.jsonRestart Windsurf for changes to take effect.
  • Claude Code.mcp.jsonSave at the project root, then start Claude Code in that project and review the MCP server approval prompt. Keep real credentials out of shared files.

Learn More