A coding agent built on Pi. One standalone executable gives you a terminal interface, a bundled web UI, Herdr integration, background jobs, sub-agents, and project wisdom.
Linux x64/arm64, macOS Apple Silicon, and Android Termux arm64:
os="$(uname -s | tr A-Z a-z)"; [ "$(uname -o 2>/dev/null)" = Android ] && os=android; asset="die-$os-$(uname -m | sed s/aarch64/arm64/ | sed s/x86_64/x64/)"; mkdir -p ~/.local/bin && cd "$(mktemp -d)" && curl -fLO "https://github.com/tnfssc/die/releases/latest/download/$asset" && curl -fLO "https://github.com/tnfssc/die/releases/latest/download/$asset.sha256" && (sha256sum -c "$asset.sha256" 2>/dev/null || shasum -a 256 -c "$asset.sha256") && install -m 755 "$asset" ~/.local/bin/diePut ~/.local/bin on your PATH. Then run:
dieTo upgrade later:
die updatedie # Start the interactive TUI
die web # Start the web UI
die -p "Describe this tree" # Run one prompt and exit
die -c # Continue the latest session
die -r # Pick a saved session to resumeInside die, type / to see available commands.
In the terminal, /questions opens the current conversation’s question inbox.
/remote opens the SSH remote-work inbox, including questions from remote tasks.
These are separate owners: a remote task’s question is answered through /remote,
not the local /questions inbox.
Type to search a menu, use arrow keys and Enter to select, and Escape to go back
without submitting. PgUp/PgDn scroll long question or choice text.
Questions offer their saved choices and, when allowed, a
custom-answer editor. Command completion also supports explicit /remote
subcommands and task/question targets; copying IDs is not required for the menus.
Remote cached state remains readable offline. Answering requires a fresh owner and question version. An uncertain reply is not a confirmed answer: reconcile the saved reply rather than replacing it with a new one. Explicit commands and the noninteractive RPC interface remain available.
Run die web and open the local URL it prints. The browser interface is built on T3 Code. It comes in the executable, so you need no separate Node or Bun install.
The bundled source is pinned to the official preview channel (v0.0.43-preview.20260921.2045), not nightly. Exact upstream revision and local integration changes are recorded in integrations/t3/upstream/source.json and integrations/t3/upstream/die.patch. See T3 integration tooling for the canonical build, gates, and historical research boundary.
Chat with die, switch models and agent modes, follow background agents, review changes, and use the integrated terminal. The server binds to 127.0.0.1 by default. Run die web --help for options.
Subagents share the current checkout by default. Work that needs isolation can use a separate Git worktree and branch. The configured setup runs there. CLI worktrees do not need the web server. See subagent workspaces for the API and retention behavior.
Run die in a Herdr-managed terminal pane and it automatically reports whether it is working, idle, or waiting for input. Background jobs and sub-agents keep the pane marked as working even after the foreground turn ends.
The integration is built in. It needs no extra extension or setup. Herdr is optional. Die also works on its own. See Herdr integration for lifecycle and compatibility details.
Install Bun 1.4.1, Node 24, and pnpm 11, then run:
bun install --frozen-lockfile
bun run check
bun run buildInstall your local build:
bun run install:localOn GitHub (including mobile): Actions → Release → Run workflow → develop → Run workflow. No version input is needed: the workflow uses latest develop, prepares the next patch version and notes unless a newer stable version is already prepared, runs the native/build/updater/licensing gates, then publishes the checked assets from the exact tested commit. Check that the Publish job succeeds; a preparation commit alone is not a release. Requires the Actions bot to have permission to push develop and tags. Failed gates leave an untagged preparation commit; see manual release recovery and limitations.
- Released binaries currently support Linux x64/arm64, macOS Apple Silicon, and Android Termux arm64.
- Credentials and model configuration are supplied at runtime, like Pi.
- State is stored under
~/.die. - See
wisdom/for durable project context and deeper feature notes.
Die limits stored job output, execute capture, and persistent session-body caching. It does not delete original session history. See resource limits for defaults, truncation semantics, storage ownership, and web shutdown behavior.
The native helper is included in macOS Apple Silicon builds. Linux Live currently requires a separately built helper; Linux releases do not bundle it.
Type /live in the local terminal to start talking; type it again to stop. Choose any supported voice with /live model (the provider follows the model); /live provider configures provider credentials without changing the voice model. Model labels show local credential readiness, not verified provider access. Gemini and OpenAI Realtime use the session’s tools directly. GPT-Live (gpt-live-1) is a voice frontend paired with your selected coding agent in the same session, retaining its model settings, tools, permissions, and history. Typed and spoken work share that coding backend. Starting Live sends audio to the selected provider and may incur API charges.
Use /live setup to check provider credentials. OpenAI Live requires a configured OpenAI API key; Codex OAuth alone is not sufficient. Keys stay out of chat. /live stop ends voice, not your agent’s jobs. Speech interruption does not cancel work; ask explicitly to stop work. GPT-Live speech transcripts remain provisional, and ambiguous requests may require clarification.
For troubleshooting, use /live status, /live mic-check, or /live speaker-check. The checks ask before opening devices and do not connect to Google. The speaker check plays a short test sound; it is not proof that speech echo or interruptions work on every route. die --live-self-test checks the embedded helper without opening devices.