sase

One developer. A team of coding agents. Tracked, reviewable, repeatable work.

sase (Structured Agentic Software Engineering, pronounced "sassy") turns Claude Code, Codex, Antigravity, Qwen Code, OpenCode, Meta's Muse Code, and xAI's Grok Build into a coordinated engineering team. One developer supervises parallel agents in isolated workspaces, with every run tracked, reviewable, and repeatable.

Status: sase is alpha software, and its interfaces and workflows are still evolving. It supports POSIX systems (Linux and macOS) only; Windows is not supported. sase assumes you have at least one supported agent CLI installed and authenticated for production work, and that you prefer opinionated git-based, workspace-per-agent workflows; if you want a standalone agent instead of a coordination layer, use those CLIs directly.

Why sase

• Launch, monitor, resume, and archive agent runs from one keyboard-driven TUI (sase's TUI).
• Run agents in parallel, each in an isolated numbered workspace clone.
• Keep prompts and multi-step workflows reusable (Macros) instead of trapped in shell history.
• Track every PR-sized unit of work with status, commits, comments, and review state (Patches).
• Keep person-owned outcomes and their status in durable Goals.
• Record project tool runs and ad-hoc commands, with their output and diagnostics, as ToolRuns.
• Schedule background and recurring agent work with the scheduler service proc.

sase does not replace coding agents; it makes agent-driven engineering dependable.

See it in action

One prompt, three live agents, three models. A single GitHub prompt fans out to Claude, Codex, and Antigravity, runs all three agents in isolated workspaces, and keeps kill controls close at hand.

Supervise every run. The Agents tab shows live status, retry chains, per-agent diffs, chats, and artifact files from one control surface.

Select the still to watch the 29-second demo.

Land tracked changes. The Patches view follows the Patch lifecycle from WIP to Submitted, with grouping, search, stitches, and diffs.

Select the still to watch the 26-second demo.

Quick start

Prerequisites: Linux or macOS (POSIX; Windows is not supported), Python 3.12+, uv, git, a text editor ($EDITOR, falling back to nvim then vim), and one authenticated agent CLI: Claude Code, Codex, Antigravity CLI (agy), Qwen Code, OpenCode, Meta's Muse Code (muse), or xAI's Grok Build (grok). SASE's default provider autodetection never selects Muse or Grok from PATH, because those executable names are generic; select either explicitly with a provider/model directive. Model-alias routing is a separate mechanism: whichever shipped size aliases currently target Muse or Grok can select them (see the generated shipped size-alias defaults). The +home examples below run in SASE's built-in home sandbox, not in your current directory.

uv tool install "sase>=0.18"                   # add a plugin too: uv tool install "sase>=0.18" --with sase-github
sase doctor                                               # check install, config, and provider authentication
sase completion install                                   # optional: <TAB> completion for zsh, bash, or fish
sase run "+home summarize what this repository does; do not change files" # auto-detected providers
# Or, for Muse Code:
sase run "%model:muse/muse-spark-1.3 +home summarize what this repository does; do not change files"
# Or, for Grok Build:
sase run "%model:grok/grok-4.7 +home summarize what this repository does; do not change files"
sase tui                                                  # open the interactive control surface

For a first run, use the auto-detected-provider command, or one of the explicit Muse/Grok examples for deterministic provider selection. The +home project tag targets the built-in home project, which is bootstrapped automatically, so the first run needs no project setup. It expands to the canonical #git:home workspace reference. After the run completes, sase tui opens the TUI with the completed run visible on the Agents tab.

If sase doctor reports a missing provider, install and authenticate it, then run the check again; see Choose providers and models. For full installation details use INSTALL.md, or follow Get started for the guided path.

Works with your agents

Agent	Status
Claude Code	Supported
Antigravity CLI (agy)	Supported
Codex	Supported
Qwen Code	Supported
OpenCode	Supported
Muse Code	Supported
Grok Build	Supported

Learn more

The complete documentation lives at sase.sh. New here? Start with Run your first agent (about 15 minutes), then pick the door that matches your goal:

• Learn — Get started — install SASE, run a first agent, and review a first change
• Do — Guides — goal-titled how-tos for setup, daily work, and fixing problems
• Understand — Concepts — how agents, work units, and landing fit together
• Look up — Reference — CLI, configuration, keymaps, and glossary

A few guides worth bookmarking:

• Add an existing repository — work in a repository you already have

• Choose providers and models — ready one provider and pick the model agents use

• Diagnose problems with sase doctor — run the read-only checks first

• Recover an interrupted agent — restart, resume, or revive a broken run

• Blog — announcements and deep dives

• The SASE Handbook (PDF) — the Guide-tier handbook and launch essay in one document

• Support — report bugs and ask questions in GitHub Issues

Development

Development requires just. Start with CONTRIBUTING.md, then see Development for the full contributor guide.

git clone https://github.com/sase-org/sase
cd sase
just install-venv
sase core health

just install-venv sets up this checkout's .venv for tests and lint — it never touches your global sase command. See Your sase versus this checkout's .venv before reaching for just install or just install-dev.

Run just check-full before submitting changes. That exhaustive local lane updates TUI screenshot goldens; CI checks them with just fix-tui-screenshots --check. SASE agents use just check unless explicitly instructed to run just check-full, and run either through sase tool run (sase tool run check or sase tool run check-full); the raw just recipes refuse to run in a SASE agent process.

Acknowledgements

sase builds on Boris Cherny's parallel-agents demonstration, Steve Yegge's beads, Agentic Software Engineering, and PDL. See the full acknowledgements.

License

sase is licensed under the MIT License. See LICENSE.