Coordinate coding agents.Review one result.
Pytxo runs the agent CLIs you already have in isolated workspaces, keeps them off each other's files, and prepares one reviewable package. In Orbit and Galaxy, prepared changes reach your repository only through explicit Apply.
Local-first. Your agent accounts and keys. Desktop ships for Windows today.

Three agents. One working tree.
Running two coding agents against one checkout is not parallelism. They overwrite each other, one reverts the other's edits mid-run, and the diff you end up reviewing belongs to neither of them.
So people serialise: one agent, wait, read the diff, next agent. The machine is idle and you are the scheduler.
Pytxo makes the claim explicit before anything runs. Tasks declare the paths they own. Overlapping claims are ordered into waves instead of racing, and each agent writes into its own isolated copy.
One screen while the work runs.
The left side is what happened. The right side is what you are about to authorise. Nothing else competes for the space.

- 01
Run ledger
Every agent task grouped by the wave that scheduled it, with the state the orchestrator actually reported and the paths the task claimed.
- 02
Reported, not predicted
A task is Completed, Running, Queued, or Failed with its exit code. There is no progress bar, because the orchestrator does not report per-agent progress.
- 03
Commit boundary
The panel that shows the prepared package, effective permission profile, and journaled Apply state. It stays visible so repository impact is never a surprise.
- 04
Enforcement receipt
Four isolation surfaces, each recorded as enforced, advisory, unavailable, or bypassed. Advisory is not a pass.
The boundary tells you what it could not prove.
Every run records how each isolation surface was actually held. A surface is enforced when the mechanism held it, advisory when Pytxo asked but the platform did not guarantee it, unavailable when nothing was measured, and bypassed when it was deliberately turned off.
Advisory and unavailable are not rendered as green. A run whose network isolation was never measured says so, on the same screen where you decide whether to apply it.
This is the whole product in one sentence: you approve an exact package, and you approve it knowing the limits of the evidence behind it.
- Workspace isolationEnforced
git-worktree
Each agent writes into its own copy of the repository.
- Host filesystemAdvisory only
child-cwd-and-policy-gates
Pytxo constrains the child process, but the OS is not enforcing it.
- NetworkUnavailable
not-reported-on-this-platform
No claim is made. Unknown is not styled as a pass or a warning.
- Apply boundaryEnforced
reviewed-run-atomic-apply
Only reviewed bytes are written, to one repository root, journalled.
Orbit and Galaxy hold prepared changes for explicit Apply. DeepSpace is non-flushable; Supernova writes directly.
It runs the agents you already installed.
Pytxo detects these CLIs by probing for their binary on your PATH. There is no Pytxo model, no provider account, and no proxy in the default path.
Anything you can run in a terminal also works through pytxo run --cmd. Editors connect through the local MCP hub instead.
One measured number, with its limits attached.
82.9%
weighted scaffold-byte reduction across 185 tracked production files. Signal Core sends syntax structure first, so an agent's first read of a file is its shape rather than its full source.
- Input
- 1,253,675 bytes
- Scaffolded
- 213,847 bytes
- Measured
- 2026-07-30
- Fallbacks
- 0
What this number is not
- Not a model-token saving. Tokenizer output was not measured.
- Not a cost saving. No provider bill was compared.
- Not a task-success claim. Agent outcome quality was not measured.
- Not a general-repository result. One repository, one commit, one host.
Questions with exact answers.
Pytxo is local-first orchestration, not another editor or opaque agent cloud.
Start with the CLI. Add Desktop when you want the boundary on screen.
The CLI does the orchestration: planning, waves, isolation, and Apply. Pytxo Desktop is an optional control surface over the same local core, so installing it never changes what a run is allowed to do.
npm i -g pytxo
What ships today
- CLI, all platformsAvailable
- Desktop, WindowsAvailable
- Desktop, macOS and LinuxNot built yet
- Managed cloud executionConfigured deployments only
Longer term, Pytxo aims to be the commit layer for autonomous work: typed effect contracts at the boundary to systems beyond a repository. That is a stated direction, not a shipped feature.
Plans and availability