- GitHub
- npm
- Claude Code
- Codex
- Cursor MCP
4.8% context on a tiny Signal fixture · aspirational ~60% on large files
More output, not more terminal panes
Many cloud agent products put demos first: lots of terminals in a browser tab, high RAM, and vendor credits. Pytxo runs agents locally in the background and records what actually changed in your code.
Your IDE stays your IDE. Agents work in isolated copies, conflicting paths get separate waves, and Cursor can drive runs through a local MCP hub.
Typical cloud agent workspace
- Walls of terminals in a heavy browser UI
- High RAM for demos
- Vendor credits and lock-in
Pytxo
- Background agents on your machine
- You pay your LLM providers directly
- Optional Desktop: change map and approvals
Up and running in three steps
No cloud signup. No terminal grid. A local agent hypervisor on your machine.
- 01Guide
Install
Install the CLI with npm, then run pytxo init and pytxo doctor in your git repo.
- 02Guide
Connect your editor
Add the local pytxo-mcp server in Cursor or VS Code to start runs from chat, or use the CLI only.
- 03Guide
Run a multi-agent job
Define agents in pytxo.toml, preview the plan, then execute in isolated worktrees.
The three moats behind every run
Signal Core, Blast Shield, and Race Shield run on every job. They are what make parallel agents safe on a real repository, not just a demo.
Less noise, lower token bills
Sends agents file skeletons (signatures, types, imports) instead of whole files, so every read starts smaller.
Safe parallel sandboxes
Each agent works in its own copy. Nothing touches your real files until you approve the merge.
Parallel runs without collisions
Dependency-aware scheduling so agents that must run in order do, without locking your whole repo.
Also included
Text-first Flow, local-first Voice, editor integration, Workspaces, fleets, and approval gates.
Mission planning with Flow
FlowDescribe a mission, review the generated tasks, permissions, ADE assignments, and execution waves, then dispatch the approved plan.
Learn moreLocal-first voice capture
VoiceTurn speech into an editable Flow mission locally. Audio stays in bounded memory and is never written to Pytxo logs or storage.
Learn moreWorks with Cursor and your IDE
MCPLocal tool bridge so you can trigger runs and inspect status from the editor you already use.
Learn moreMulti-folder Workspaces
Multi-rootOne coordinated run across API, web, and shared folders. Each folder can have its own trust level.
Learn moreCross-repo fleet runs
FleetCoordinate agents across separate git repos when steps must finish in order, with status in Pytxo Desktop.
Learn moreApproval gates for risky actions
ApprovalsRequire a human OK before sensitive spawns or applying batched changes, from CLI, Desktop, or your editor.
Learn moreWorks with your CLI
Pytxo coordinates the agent. You choose the CLI. Run Antigravity (agy), Claude Code, Codex, or any shell command via pytxo run --cmd.
| Agent | How it runs | Mode |
|---|---|---|
| Antigravity CLI (agy) | Background agent process | Terminal agent |
| Claude Code | Background agent process | Your API keys |
| OpenAI Codex CLI | pytxo run --cmd "codex …" | Your API keys |
| Aider / custom scripts | Any shell command | Your API keys |
| Cursor | Editor integration (MCP) | Local tools |
Local control plane, not a browser full of terminals
Pytxo plugs into your existing workflow. It coordinates agents and records what changed, so Desktop is a branch off the same pipeline, not a separate product.
Editor or terminal
Cursor, VS Code, or shell
MCP hub
Optional local IDE bridge
Orchestration
Rust: projects, permissions, waves
Execution yard
Claude Code, Codex, and other CLIs
Pytxo Desktop
OptionalStructural telemetry and approvals
Install once. Run agents in parallel.
CLI for orchestration. Optional Desktop for topology, approvals, and diffs.
npm i -g pytxo