Hypa

Hypa is a workspace for coding agents. It runs your agents in terminal panes, keeps them alive when you close the terminal, lets agents drive other panes, and cuts the noise from command output before it reaches an agent.

Hypa is local-first. It needs no cloud service.

What Hypa does

• Workspace mux. hypa starts a local server that owns your panes. Attach from any terminal, detach, and attach again. Tabs, splits, workspaces, a sidebar, copy mode, popups, themes, and settings are built in.
• Agent runtime. An agent in a Hypa pane can read other panes, start work in a new pane or a hidden pane, wait for another agent, and report its state. Hypa installs a skill that teaches an agent how.
• Share and connect. Share a session with another machine. The other machine connects over QUIC, with TCP and TLS as the fallback. A certificate pin protects every connection.
• Output compression. hypa -c "command" runs a command, removes noise from the output, and keeps what matters: errors, warnings, file paths, failing tests, and exit codes. The reduction is deterministic and local. It is not an LLM summary.

Install

Pick the channel that fits how you manage tools. Every channel installs the same release archive for your platform, so the mux works the same way in each.

Releases from 1.0.1 on are published to Homebrew, npm, and PyPI. Version 1.0.0 is available only as a release archive and through the install script.

Install one copy. If you switch channels, remove the old copy first, so that only one hypa is on your PATH.

Check any install:

hypa --version
hypa doctor

Requirements

The npm package needs Node.js 18 or newer. The PyPI package needs Python 3.9 or newer. Neither needs a .NET runtime.

Install script

Linux and macOS:

curl -fsSL https://raw.githubusercontent.com/Hypabolic/Hypa/main/install.sh | sh

The installer downloads the release archive for your platform, checks it against SHA256SUMS, and installs hypa in a bin directory that you can write to (~/.local/bin by default). It prints a warning when that directory is not on your PATH.

Installer options:

HYPA_VERSION=1.0.1 HYPA_INSTALL_DIR="$HOME/bin" sh install.sh
HYPA_ARCHIVE=/path/to/hypa-osx-arm64.tar.gz HYPA_INSTALL_DIR="$HOME/bin" HYPA_APP_DIR="$HOME/share/hypa" sh install.sh

Homebrew

macOS and Linux:

brew install hypabolic/tap/hypa

This adds the hypabolic/tap tap and installs the formula. Homebrew keeps the whole release directory in the Cellar and links hypa into your PATH.

npm

npm install --global @hypabolic/hypa

The @hypabolic/hypa package picks the matching platform package, for example @hypabolic/hypa-darwin-arm64, as an optional dependency. If your package manager skips optional dependencies, install the platform package yourself.

PyPI

pipx install hypa
# or
uv tool install hypa
# or
pip install hypa

Prefer pipx or uv tool, which keep hypa out of your project environments. The wheel carries the native build, and the hypa command starts it.

Release archive

Download the archive for your platform from GitHub Releases, together with SHA256SUMS, and check the archive against it.

The archive unpacks to a hypa-<platform> directory. Keep its files together: hypa finds the mux host, the attach client, and its native libraries next to itself. Run hypa from that directory, or link it into a directory on your PATH.

Update

hypa update --check shows whether a newer release exists. hypa update upgrades a copy that the install script put in place. For a copy that a package manager installed, it prints the command to run:

After an upgrade, restart a running mux server (hypa mux stop, then hypa attach) so that it runs the new version.

Windows

Windows is not a mux host. There is no Windows mux archive, and install.ps1 stops with an error. The compression CLI can still run on Windows. Build it from source (see below).

Quick start

hypa

Bare hypa starts the local mux server, or reconnects to it, and attaches. The first run shows an onboarding overlay.

Keys use a prefix, ctrl+b by default:

Close the terminal, or detach, and attach again with hypa. Panes keep running while you are away.

Install the agent integration. Hypa finds the agents on your machine and shows every file it writes before it writes:

hypa integration install
hypa integration status

hypa init installs hooks and skills for the detected agent harnesses. Use --dry-run to preview either command.

Workspace mux

A session holds workspaces, each workspace holds tabs, and each tab holds panes. One mux server owns one session.

hypa --session NAME                  # attach to a named session
hypa session list
hypa session attach NAME
hypa session stop NAME
hypa session delete NAME
hypa status                          # mux client and server status
hypa mux serve --session default     # run the server without attaching
hypa mux stop

hypa ping, hypa snapshot, hypa workspace, hypa tab, hypa pane, and hypa layout talk to a running server. They do not start it.

Attach to a mux on another machine through OpenSSH:

hypa --remote user@host
hypa --remote user@host --session NAME

OpenSSH owns keys and credentials. Hypa stores none. See docs/guides/remote-ssh-attach.md.

Panes use a real PTY and the Ghostty terminal engine. Hypa paints one composed frame for your terminal. The host palette, focus events, mouse, and bracketed paste reach the panes, and arrow, Home, and End keys use the form that the running program asks for.

Customize Hypa in ~/.config/hypa/config.toml:

hypa --default-config                # print the default file
hypa config check                    # validate your file

There are 19 built-in themes. An invalid value in the file stops Hypa with an error. It does not fall back in silence.

A saved screen history for restored panes is off by default. See docs/guides/session-state.md.

Agent runtime

Every pane child gets HYPA_ENV=1, HYPA_BIN_PATH, HYPA_RUNTIME_SOCKET, and pane identity keys. An agent uses the hypa command with those keys to control the session. See docs/guides/pane-environment.md.

hypa --skill                         # print the agent skill
hypa pane --help                     # pane operations for scripts and agents
hypa api schema --json               # the control-plane protocol

What an agent can do:

• Inspect workspaces, tabs, panes, and agents.
• Split a pane and run a command, then read its output or wait for it.
• Start another agent in a new pane.
• Create a hidden pane for background work. A hidden pane has a live PTY and no place in the layout. You can show, hide, and read it from the attach view. See docs/guides/background-panes.md.

The skill is installed for the agents that Hypa detects. See docs/guides/agent-skill.md. Hypa supports many agent harnesses, including Claude, Codex, Copilot, Cursor, Grok, Pi, and others. Run hypa integration install --help for the list.

A plugin runs as your user with full permissions. It is not a sandbox.

Share and connect

Open the share dialog in the attach view to make an invite. The invite carries every reachable address of your machine, so you do not set an address first. The other machine redeems the invite, tries each address, and reports the address that answered.

• QUIC is the first path. TCP with TLS is the fallback.
• A certificate pin protects every connection. The secret is sent only after the pin matches.
• A saved peer is a cube. Connect moves your view to that peer.
• hypa connectivity accept runs the listening side by hand. hypa device pairs and revokes devices.

An invite holds at most 32 addresses. Use --advertise-host to name one address. See docs/guides/share-invite.md and docs/guides/connectivity-accept.md.

Output compression

Hypa runs a command, reduces the output, and returns a compact result:

shell command
  -> Hypa command runner
  -> command-specific reducer, when available
  -> built-in or trusted DSL filter, when applicable
  -> token accounting and savings metadata
  -> compact output returned to the caller

If the result saves tokens, Hypa appends a footer:

[hypa: 1200→340 tok, -72%, reducer=dotnet-build]

For failures or truncation, Hypa can tee the full output to a local artifact, so the compact output stays small and you can still recover everything.

hypa -c "dotnet test"                # run and compress
hypa -t dotnet test                  # run unmodified, streamed to the terminal
hypa git status                      # first-class wrappers: git, dotnet, kubectl, docker
hypa rewrite "git status"            # show how a command would be rewritten
hypa filters list                    # built-in and configured filters
hypa filters test NAME ./output.txt  # test a filter against a saved output
hypa filters savings --markdown      # estimate savings for the filter suite
hypa read PATH                       # read a file in a context-aware mode
hypa search QUERY                    # search files, symbols, and indexed context
hypa code                            # index and query source code structure
hypa compress                        # compress text from stdin or a file

Built-in filters cover build and test tools (dotnet, cargo, gradle, mvn, go test, jest, pytest, xcodebuild), package managers (npm, pnpm, yarn, pip, poetry, uv), linters, infrastructure tools (terraform, helm, kubectl, docker, aws, gcloud), system tools, monorepo task runners, and source control. Run hypa filters list for the current list.

Token estimates use Microsoft.ML.Tokenizers with the o200k_base tokenizer. Savings reports use synthetic payloads. Treat them as repeatable estimates, not as a measure of your own projects.

Hypa runs trusted project filters from a repository .hypa/ directory. It does not run every repository filter on its own. Use hypa trust status and hypa trust filters.

Hypa also works as an MCP server (hypa serve) and through agent hooks (hypa hook, installed by hypa init). The Pi extension has its own guide: docs/guides/pi.md.

Local data

Hypa keeps runtime data in ~/.hypa/:

• hypa.db: SQLite database with sessions, command metrics, parse metrics, trust records, and artifacts.
• artifacts/: full command output, kept for recovery.
• config.json: optional compression configuration.

The mux configuration is ~/.config/hypa/config.toml. It is a separate file from config.json. Do not merge them.

Platforms and limits

• Linux and macOS are the mux hosts (linux-x64, linux-arm64, osx-x64, osx-arm64).
• Windows is not a mux host. See docs/ADRs/0015-windows-is-not-a-mux-host.md.
• A running mux keeps the binary that started it. After an upgrade, stop the old mux and start it again.
• QUIC does not start when you bind hypa connectivity accept to a specific address. That accept uses TCP with TLS.

Build from source

You need the .NET 10 SDK. Linux or macOS is required for the mux.

dotnet build
dotnet test
dotnet format --verify-no-changes --no-restore
dotnet run --project src/Hypa.Cli -- --help

Publish a local binary (Linux x64 example):

dotnet publish src/Hypa.Cli/Hypa.Cli.csproj -c Release -r linux-x64

The executable is written under src/Hypa.Cli/bin/Release/net10.0/linux-x64/publish/. Put that directory on your PATH, or link the executable as hypa.

To run the current working tree as your hypa command while a release is installed, use the preview helper:

scripts/hypa-preview install --test    # publish, test, and switch to the preview
scripts/hypa-preview install --fast    # skip tests and NativeAOT
scripts/hypa-preview status
scripts/hypa-preview restore           # switch back to the release

While a preview is active, hypa update does not replace it.

Documentation

• docs/guides/: how-to guides for the mux, the agent skill, panes, sharing, and remote attach.
• docs/architecture/ and docs/ADRs/: design and decisions.
• docs/release-notes/: release notes.

Status

Hypa 1.0.0 is the first release with the workspace mux and the agent runtime. Report faults at https://github.com/Hypabolic/Hypa/issues.

License

Hypa is licensed under the Functional Source License, Version 1.1, with an Apache 2.0 future license. See license.md in the public repository.