让 Linux 上的 Claude Desktop 能够启动、监控和控制 tmux 会话中长时间运行的任务。专为机器学习训练、数据流水线以及任何你希望在单次对话结束后依然保持运行的任务而设计。
Claude 可以启动一个后台训练任务，一小时后通过捕获其窗格的最后 N 行来检查进度，发送按键（包括 Ctrl+C），并在完成后终止会话。你可以从任何真实终端执行 tmux attach -t <name> 来实时观看同一个会话——Claude 和人类可以同时观察。
Linux 且已安装 tmux：
sudo apt install tmux适用于默认的每用户 tmux 服务器（无需共享服务器设置）。
Related MCP server: Tmux MCP Server
从 Releases 页面下载最新的 Sessions.mcpb。
在 Claude Desktop 中：Settings → Extensions → Extension Developer → Install Extension → 选择 .mcpb 文件。
启用该扩展。
工具 | 注解 | 功能 |
| 只读 | 列出所有 tmux 会话，包含名称、窗口、创建/活动时间戳、附加标志 |
| 只读 | 精确匹配检查会话名称 |
| 只读 | 会话内的窗口（索引、名称、活动状态、窗格数量） |
| 只读 | 会话活动窗格的最后 N 行（默认 100，最大 10000） |
| 破坏性 | 启动一个后台会话；可选择运行初始 shell 命令 |
| 破坏性 | 向活动窗格发送按键组合或文本（默认自动追加 Enter） |
| 破坏性 | 终止会话（向内部所有进程发送 SIGHUP） |
启动训练任务：
“启动训练任务
b8-c1：激活我的rdna4-train虚拟环境并运行 train.py。”
new_session name=b8-c1 command="source ~/.venvs/rdna4-train/bin/activate && python /media/luke/Storage\ Device\ 1/NeuroForge/train.py b8-c1"
稍后检查：
“训练
b8-c1还在运行吗？给我看最后 50 行。”
capture_pane name=b8-c1 lines=50
中断卡死的任务：
“训练卡住了——发送 Ctrl+C。”
send_keys name=b8-c1 keys="C-c" enter=false
结束任务：
“终止会话。”
kill_session name=b8-c1
send_keys 接受字面文本（作为按键发送）或 tmux 按键名称。常见按键：
按键 | 效果 |
| Ctrl+C (中断) |
| Ctrl+D (EOF / 登出) |
| Escape 键 |
| 方向键 |
| 回车键（当 |
| 空格键 |
发送原始组合键时请使用 enter: false，以免额外触发回车。
此扩展完全在你的本地机器上运行，仅调用 tmux 二进制文件。没有任何数据离开你的机器。没有网络 I/O。没有套接字。除了 tmux 自身的每用户服务器套接字（默认 /tmp/tmux-$UID/default）外，不会写入任何文件。
Claude 可见的信息包括会话名称、窗格内容（你或你的程序打印的文本）以及会话元数据（创建/活动时间戳、窗口/窗格数量）。send_keys 和 kill_session 可能会影响在受管会话中运行的任何进程，因此请给予它与你信任用户在终端输入时相同的信任级别。
“tmux is not installed” — 运行 sudo apt install tmux。
“can't find pane” 错误 — 通常意味着命名会话在操作过程中被终止了。请先调用 has_session，或使用 list_sessions 查看当前实际运行的内容。
会话名称限制 — 名称不能包含 .、:、| 或空格。. 和 : 是 tmux 目标分隔符；| 是此服务器的字段分隔符；空格原则上被拒绝。
会话在初始命令后无法存活 — 根据设计，new_session 总是启动一个 shell；command 参数作为按键发送到该 shell 中。这就是为什么命令结束后会话依然保持存活的原因。如果你直接将 command 传递给 tmux new-session，它会在命令结束时退出。
单个约 300 行的 Node.js 脚本，零 npm 依赖。重新构建 .mcpb：
cd bundle-source
zip -j ../Sessions.mcpb manifest.json package.json server.js README.md LICENSE icon.png glama.jsonMIT。参见 LICENSE。
claude-terminal-mcp — shell、文件系统和后台任务（适用于 tmux 显得过于笨重的短时命令）。
claude-linux-mcp — X11 桌面控制。
claude-rocm-mcp — AMD GPU 监控；与此工具配合使用可在训练期间检查 GPU 健康状况。
Capture the last N lines visible in the session's active pane (default 100, max 10000). Returns the joined text. The primary primitive for "is my training still running? show me the latest output".
| Name | Required | Description | Default |
|---|---|---|---|
| name | Yes | Session name. | |
| lines | No | Number of lines to capture (default 100). |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
Annotations already indicate readOnly and non-destructive behavior. The description adds that it captures from the 'active pane' and returns 'joined text', which is useful behavioral context beyond annotations.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
Two sentences: first technical with defaults/limits, second a relatable use case. No wasted words, information is front-loaded and clear.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
For a simple read-only tool with two parameters and no output schema, the description covers purpose, behavior, parameters, and use case fully. No gaps.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Input schema has 100% coverage with descriptions for both parameters. The description adds the default value for 'lines' (100), which is not in the schema, giving minor additional value. Baseline 3 is appropriate.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description clearly states the action ('Capture the last N lines'), the resource ('session's active pane'), defaults and limits, and a concrete use case. It distinguishes from sibling tools which manage sessions or send keys, not capture output.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description provides a clear context for use ('is my training still running?') but does not explicitly state when not to use or name alternatives. However, the sibling tools imply that this is the sole tool for capturing output.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Check whether a named tmux session exists. Cheap precondition for capture_pane, send_keys, or kill_session.
| Name | Required | Description | Default |
|---|---|---|---|
| name | Yes | Session name. Cannot contain ".", ":", "|", or whitespace. |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
Annotations already indicate readOnlyHint=true and destructiveHint=false. Description adds that the operation is 'cheap', which is valuable behavioral context beyond annotations.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
Two sentences, no wasted words, front-loaded with purpose. Efficient and clear.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
Given the simplicity of the tool (one parameter, no output schema), the description is complete: it explains purpose, usage context, and constraint hints.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema coverage is 100% with the parameter name described and constraints noted. Description does not add additional semantic meaning beyond what the schema provides, so baseline 3 is appropriate.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
Description clearly states the tool checks if a tmux session exists, using a specific verb and resource. It distinguishes itself from sibling tools by mentioning it's a precondition for three specific actions.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
Description explicitly states it's a cheap precondition for capture_pane, send_keys, or kill_session, providing clear usage context. No exclusions or alternatives are mentioned, but the context is sufficient.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Terminate a named tmux session. Any processes running inside are sent SIGHUP by tmux. Errors if the session does not exist.
| Name | Required | Description | Default |
|---|---|---|---|
| name | Yes | Session name. |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
Adds beyond annotations: details SIGHUP behavior and error condition if session does not exist. Matches destructiveHint.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
Two concise sentences, front-loaded with primary purpose. No wasted words.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
Complete for a simple kill tool with one parameter and no output schema. Covers purpose, behavior, and error case.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema has 100% coverage but only a minimal description for the 'name' parameter. No additional semantics added in tool description.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
Clearly states 'Terminate a named tmux session' with a specific verb-resource pair. Distinct from sibling tools like capture_pane or list_sessions.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
Conveys when to use (to terminate a session) and mentions consequences (SIGHUP, error if missing). Could note alternatives but sufficient.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
List all tmux sessions on the current user's tmux server. Returns name, window count, created/activity timestamps (ISO 8601), and attached status. Returns an empty list when no tmux server is running.
| Name | Required | Description | Default |
|---|---|---|---|
No parameters | |||
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
Annotations already provide readOnlyHint and destructiveHint. Description adds beyond that by detailing return fields (ISO 8601 timestamps, attached status) and the edge case of an empty list when no tmux server is running.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
Two sentences, front-loaded with the main action, and every word adds value. No fluff or redundancy.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
Given no parameters, no output schema, and annotations, the description fully covers what the tool does, its return values, and a key edge case (no server running).
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
No parameters exist, and schema coverage is 100% (trivially). Per guidelines for 0 parameters, baseline is 4. Description does not need to add parameter details.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
Description clearly states 'List all tmux sessions' with specific return fields (name, window count, timestamps, attached status). It distinguishes from sibling tools like kill_session and new_session by focusing on listing only.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description implies usage for reading sessions, not modifying them. However, it does not explicitly state when not to use or mention alternatives. The context from sibling tools helps, but clarity could be improved.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
List windows within a named session (index, name, active flag, pane count).
| Name | Required | Description | Default |
|---|---|---|---|
| name | Yes | Session name. |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
Annotations already declare readOnlyHint=true and destructiveHint=false, so the description adds the return fields (index, name, active flag, pane count). However, it does not disclose behavior if the session does not exist or any other edge cases. Adds some context but not rich behavioral details.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is a single, concise sentence (12 words) with no wasted words. It is front-loaded with the action 'List windows' and immediately clarifies scope and return fields.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
For a simple read-only list tool with 1 parameter and no output schema, the description is fairly complete: it specifies what it does, what it returns, and the required input. It could mention that the session must exist, but that is implied. Sibling tools provide additional context.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema coverage is 100% (the only parameter 'name' is described as 'Session name'). The description reinforces that the session is named but adds no new semantic meaning beyond the schema. Baseline 3 for high coverage.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description clearly states the verb 'List' and the resource 'windows within a named session', and specifies return fields (index, name, active flag, pane count). This distinguishes it from sibling tools like list_sessions (list sessions) and capture_pane (capture pane output).
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description implies usage when you have a session name and want to see its windows, but does not explicitly state when to use this tool versus alternatives like list_sessions or when not to use it. No guidance on prerequisites (e.g., session must exist).
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Start a new detached tmux session. Always starts a shell (so the session survives after the initial command finishes). If command is given, it is sent to the shell as keystrokes + Enter, so aliases, venvs, and functions work as expected. Errors if the session name already exists.
| Name | Required | Description | Default |
|---|---|---|---|
| name | Yes | Session name (unique). Cannot contain ".", ":", "|", or whitespace. | |
| command | No | Optional initial command to run inside the shell (e.g. "cd /path && python train.py"). | |
| cwd | No | Optional starting working directory for the session. |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
Adds significant behavioral details beyond annotations: detached, always starts shell, command sent as keystrokes enabling aliases, errors on duplicate name. No contradictions.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
Three sentences with no waste. First sentence states purpose, second explains key behavior, third clarifies command detail and error condition. Front-loaded and efficient.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
Covers purpose, behavior, error conditions, and parameter semantics. No output schema needed. Complete for a creation tool with 3 parameters.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema coverage is 100%, so baseline is 3. Description adds key nuance: command is sent as keystrokes + Enter, preserving aliases and venvs. This raises the score.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description clearly states the tool starts a new detached tmux session, which is a specific verb + resource. It distinguishes from siblings like 'kill_session' by focusing on creation.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
Describes when to use (to start a session) and conditions: always starts a shell, errors if name exists. No explicit comparison to alternatives, but context is clear.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Send key combos or text to the session's active pane. Use for interrupts (C-c), end-of-input (C-d), or typing a command. Text is sent literally. Enter is appended by default; set enter=false for raw key combos.
| Name | Required | Description | Default |
|---|---|---|---|
| name | Yes | Session name. | |
| keys | Yes | Key combo (e.g. "C-c", "C-d", "Escape") or literal text to type. | |
| enter | No | Append an Enter keypress after the keys. Default true. |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
Annotations indicate destructive and open-world hints; the description adds detail: text is sent literally, Enter appended by default, and raw combos require enter=false. No contradictions.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
Three concise sentences, front-loaded with purpose, then specific guidelines. No wasted words.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
Covers key behavioral aspects; lacks description of return values, but given no output schema and the nature of the action (send), this is acceptable. Sibling tools are listed for context.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema covers 100% of parameters; description enriches meaning by explaining literal sending and enter default behavior, going beyond what the schema provides.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description clearly states the verb (send), resource (key combos or text to session's active pane), and specific use cases (interrupts, end-of-input, typing a command). It distinguishes from sibling tools like capture_pane or kill_session.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
Provides clear when-to-use guidance (interrupts, end-of-input, typing) and explains the enter parameter behavior. Does not explicitly mention when not to use or alternatives among siblings, but sibling tools are sufficiently different.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
v0.1.0capture_panehas_sessionkill_sessionlist_sessionslist_windowsnew_sessionsend_keysScored across 7 tools
Each tool has a clearly distinct purpose: creation, output capture, listing sessions/windows, existence check, input sending, and termination. No two tools overlap in functionality, making misselection unlikely.
All tools follow a consistent verb_noun pattern in lowercase snake_case (e.g., new_session, capture_pane, list_sessions). The naming is uniform and predictable, with no stylistic deviations.
7 tools is a well-scoped set for tmux session management, covering the essential lifecycle operations without redundancy or bloat. Each tool serves a clear purpose.
The surface covers the core lifecycle (create, read, check, update via send_keys, delete) adequately. Minor gaps exist, such as no rename or attach/detach operations, but these are not critical for typical usage and can be worked around.
Work on your own computers from Claude or ChatGPT: run commands, edit and search files.
Use your Mac, Windows or Linux computer from ChatGPT, Claude or Codex: files, commands, documents.
Run UX research from Claude — create card sort studies, list studies, pull headline stats.
I run shell commands on your private cloud environment (bash, sh, zsh)
