A terminal companion that rides along in your shell session, building context as you work. When you need help, it's already up to speed.
Upstream timeout from nginx. worker_processes is likely saturated. Check:
You're the one doing the work. You're in the terminal, SSH'd into a box, tailing logs, running commands, figuring things out. When you want a second opinion from AI, you have to stop, copy output, explain what you've been doing, and rebuild the context from scratch. Every single time.
Tether fixes that. Run your shell through tether shell and it rides along in the background,
capturing output as you work. When you ask a question in tether chat,
Claude already knows what's been happening. No copy-paste. No re-explaining. No context loss.
SSH into servers, read logs, run commands. Tether is invisible until you need it. No interruptions, no approval dialogs breaking your flow.
Everything in your session is captured in real time and held in memory: commands, output, errors. Nothing is sent anywhere until you ask. Nothing is written to disk.
Open tether chat in any terminal window and ask. Claude already knows what's going on. No copy-paste, no re-explaining, no context loss.
Tether doesn't just dump your terminal history into every message. It watches as you work, then figures out what's actually relevant to your question, so Claude has the right picture without you spending a fortune on tokens or doing any curation yourself.
/debug inside chat to see exactly what Tether included in each message and why.
Switch between modes at any time with tether mode <name>. Start conservative and open up as context warrants.
Read-only. No execution risk.
Claude can answer questions, explain errors, and suggest commands in plain text. It cannot type into your terminal. The default: no risk, all insight.
Propose + approve. You stay in control.
Claude proposes commands. A proposal UI appears: you approve, edit, or reject each one with a single keystroke. Press [t] in the chat TUI to enable auto-run, which automatically executes commands on your allow list.
$ tether mode watch # read-only, no execution
$ tether mode assist # proposals, you approve each oneYou've been in control, watching, asking questions, narrowing things down. When you know what needs doing and want Claude Code to take over, type /handoff <task>. Tether bundles everything it knows into a structured brief: your current directory, git state, detected errors, and the relevant terminal output. Claude Code launches with it as its first message and starts informed, not cold.
Claude sees your terminal history before you ask anything. No copy-paste, no re-explaining what you've been doing.
Only new output is sent each message, not the full buffer. You're not paying for 500 lines of log output every time you ask a question.
Context is scored against your question. Irrelevant lines are dropped. The 5,000-line buffer is trimmed to up to 200 focused lines before anything is sent.
Tether captures everything in your local PTY, including SSH session output. No install needed on remote hosts. Works anywhere you can SSH.
Ctrl+K stops a streaming response mid-flight. Change your mind, rephrase, and ask again. No need to wait for Claude to finish.
Chat history persists across sessions and compacts automatically. Pick up right where you left off, without token bloat.
A background summarizer generates a narrative of your session every 5 minutes. Claude gets the big picture without raw line-by-line history.
Each response shows estimated prompt and response token cost. Type /debug in chat for a full breakdown: keywords extracted, lines fetched vs. filtered, and per-component token estimates.
Every command Claude suggests passes through a classifier before anything happens. Claude can influence what is proposed, but never what is executed.
curl | bash, wget | sh). Always blocked. Not overridable by anyone.sudo, output redirects (>, >>, tee), command chaining (&&, ;). Always proposed, never auto-executed.deny, protect, and allow lists in ~/.tether/config.json. Full control over what is trusted.[a] in the proposal UI. It auto-executes for the rest of the session without touching your config.
The installer detects your OS and architecture, downloads the right binary to /usr/local/bin,
and sets up shell integration automatically. The only prerequisite is the
Claude Code CLI: claude must be on PATH and logged in.
Tether uses it for all AI calls, so there's no separate API key or billing.
curl -fsSL https://raw.githubusercontent.com/AllDayJon/Tether/main/install.sh | shtether shell
# work normally — tether watches in the backgroundtether chat
# ask anything — Claude already has contextclaude must be on PATH and logged intether installInstall shell integration (bash/zsh/fish)tether uninstallRemove shell integration and configtether doctorCheck all dependenciestether shellStart your shell with context capturetether statusActive sessions, mode, daemon statetether logsStream the daemon log livetether chatOpen the chat TUI in any terminaltether ask <question>One-shot question to stdouttether mode [watch|assist]Show or change modetether tokensContext buffer sizetether summaryRolling session summarytether historyConversation history + token statsThe core difference is who's doing the work. With most AI tools, you hand the AI a goal and it executes. That's great for planned tasks, but it only knows what you tell it. The moment you go off-script and start doing things yourself, the AI loses context entirely.
Tether is built for when you're already at the wheel. You work normally: SSH into servers, tail logs, poke around. Tether keeps Claude up to speed in the background. When you want help, you just ask. No copy-paste, no re-explaining what you've been doing for the last ten minutes.
Tether itself is free and open source. The only requirement is the Claude Code CLI. You'll need a Claude subscription with claude installed and logged in on your machine. Tether uses it for all AI calls, so there's no separate API key, no extra billing, and no Tether account.
A focused snapshot of your recent terminal activity, filtered for relevance to your question, plus your conversation history and a brief summary of your session. Tether is careful about what it includes; irrelevant output is dropped so you're not burning tokens on noise. Type /debug in chat to see exactly what went into any message.
No. Tether has no servers of its own and collects no analytics. The only network traffic is the Claude Code CLI (claude -p) talking to Anthropic's API, the same traffic as using Claude Code directly, governed by Anthropic's privacy policy.
No. tether shell runs on your local machine and captures everything you see, including output from SSH sessions running inside it. Tether never needs to be on the remote host.
Whatever you already use. Run tether shell to start your session and tether chat in a second terminal window or pane. If you use tmux, iTerm2, Kitty, or Alacritty, it all works the same way.