PiloTY
PiloTY is an MCP server that provides AI agents with a persistent, interactive terminal, enabling long-running commands, log monitoring, and stateful shell interactions across tool calls.
README
<p align="center"> <img src="assets/logo.png" alt="PiloTY logo" width="220" /> </p>
PiloTY
PiloTY (PTY for your AI Copilot) is an MCP server that gives an agent a persistent, interactive terminal.
If you have used Claude Code / Codex to run shell commands, you have probably hit the same wall: tool calls tend to be stateless. Each call starts "fresh", so environment variables disappear, interactive programs cannot be driven reliably, and long-running processes get cut off or orphaned while the agent is thinking.
PiloTY exists to make the agent's terminal behave more like a human's: start something in a real terminal, come back later, and keep going.
Warning: PiloTY exposes unrestricted terminal access. Treat it like giving the agent your keyboard.
What it enables
- Long-running commands: builds, installs, migrations, test suites. Start once, check output later.
- Log monitoring:
tail -f,journalctl -f,kubectl logs -f, CI logs, service restarts. - "Vibe debugging": keep a REPL/debugger open while the agent reads code and tries ideas (
python,ipython,pdb). - Privileged operations: handle interactive password prompts (
sudo, SSH passwords, key passphrases). - SSH-based devops: keep a remote login session alive across tool calls; run remote commands in the same shell.
- Terminal UIs:
less,man,top,vimcan work, but cursor-heavy programs often require screen snapshots instead of plain text output.
Quickstart
PiloTY is meant to be launched by an MCP client over stdio.
Add it to Codex CLI as an MCP server:
codex mcp add piloty -- uvx --from git+https://github.com/yiwenlu66/PiloTY.git piloty
If you prefer SSH-based Git fetch:
codex mcp add piloty -- uvx --from git+ssh://git@github.com/yiwenlu66/PiloTY.git piloty
If you already have a local clone:
codex mcp add piloty -- uv --directory /path/to/PiloTY run piloty
Run the server command directly (without adding it to an MCP client):
uvx --from git+https://github.com/yiwenlu66/PiloTY.git piloty
Mental model
One session is one real interactive terminal that stays alive across tool calls.
- State persists: cwd, environment variables, foreground process, remote SSH connection, REPL/debugger state.
- PiloTY tracks both the raw output stream and a rendered screen/scrollback view.
PiloTY keeps two representations:
output: incremental text stream (optionally ANSI-stripped)- Rendered screen/scrollback: what a human would see in a terminal
Public MCP results separate call outcome from terminal interpretation:
outcome:success,deadline_exceeded,eof,error,invalid_session, orterminatedterminal_state: best-effort rendered-state classification after the call (running,ready,password,confirm,repl,editor,pager,unknown)
Sessions are addressed by a session_id string. Reusing the same id is what keeps state.
Integration notes (for MCP integrators)
MCP does not expose a standard "client cwd" field, so the first step is always to create a session with an explicit working directory, then reuse the same session_id for subsequent calls.
Typical agent workflow:
- Create a session (explicit cwd) and reuse the same
session_id. - Use
send_lineto submit a newline-terminated command,send_textfor raw bytes, andsend_control/send_signalfor interrupts. - Use
wait_for_outputfor temporal PTY output waiting,wait_for_regexfor content-based waits, andwait_for_shell_promptaftersshor similar login flows. - Use
snapshot_screen/snapshot_scrollbackwhen layout matters. Snapshot tools are passive and do not ingest fresh PTY bytes. - If prompt detection is wrong, configure a custom shell-prompt regex.
- Use
send_passwordfor secret entry; terminate the session when done.
For exact tool names, arguments, and return fields, use your MCP client's tool schema or read piloty/mcp_server.py.
Limitations
deadline_sis a wall-clock budget. On send/wait tools, it is not "process completion time".- Drain-based tools (
send_line,send_text,send_control,send_password,send_signal,wait_for_output) stop after the server quiescence policy (PILOTY_QUIESCENCE_MS, default1000) or whendeadline_sexpires. wait_for_outputcan return partial output withoutcome=deadline_exceededif output started but the PTY never went quiet before the deadline.wait_for_regexfirst checks already-rendered scrollback, then waits on new PTY bytes.wait_for_shell_promptconsumes PTY output while it waits and returns the consumed bytes inoutput.- Terminal-state detection is best-effort and can be wrong (especially for custom prompts and cursor-heavy TUIs).
- Plain text output can be misleading for full-screen programs; use screen snapshots when layout matters.
send_password()suppresses transcript logging and terminal echo for that send. It does not prevent other prompts/programs from echoing secrets later.- Quiescence-based output collection can be confused by programs that print periodic noise. Tune with
PILOTY_QUIESCENCE_MS(default1000).
Logs
Each server instance writes session logs under ~/.piloty/:
~/.piloty/servers/<server-instance-id>/sessions/<session-id>/transcript.log: raw PTY bytes (combined stdout/stderr)~/.piloty/servers/<server-instance-id>/sessions/<session-id>/commands.log: inputs sent (best-effort)~/.piloty/servers/<server-instance-id>/sessions/<session-id>/interaction.log: inputs plus captured output (best-effort)~/.piloty/servers/<server-instance-id>/sessions/<session-id>/session.json: metadata snapshot~/.piloty/active/<server-instance-id>/<session-id>: symlink to the current session directory (when symlinks are supported)
Server logs default to /tmp/piloty.log.
tools/session_viewer.py can inspect sessions:
python tools/session_viewer.py list
python tools/session_viewer.py info <server-instance-id>/<session-id>
python tools/session_viewer.py tail -f <server-instance-id>/<session-id>
Development
Repository layout:
piloty/
core.py # PTY + terminal renderer + session logs
mcp_server.py # MCP tools + state inference
tests/
tools/
pty_playground.py
session_viewer.py
Run tests:
python -m pytest -q
License: Apache License 2.0, see LICENSE.
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。