Start here
Orcyx is a desktop agentic development environment built from a handful of named parts: Seats (a git worktree per agent pane), the Gate (the advisory auditor scan a swarm's finished work gets before it reaches you), the Ledger (the per-workspace requirements record every pane and swarm writes to), the Vault (the markdown knowledge graph on your disk, served over MCP), Time travel (shadow-repository checkpoints of a workspace with diff and restore), Sage (voice, on your own machine and your own subscriptions), and an MCP server other tools can drive. This documentation describes what the binary does today, at the commit named on each page.
Checked against the Orcyx source tree at c339cf993 on 2026-09-14. Where this page describes something the build does not do yet, it says so in place rather than describing the intention.
- The ten-minute first run - the recipe, actually executed and timed on one machine, with the parts that could not be executed named.
- Install - per OS, and the unsigned-binary prompts you will actually see.
- Seats and worktrees - how a pane gets its own checkout, and the traps in sharing one.
- Swarm and the review gate - roles, dispatch, and exactly what the auditor gate does and does not stop.
- Build end to end - describe an outcome and Orcyx plans, builds, runs the checks you accepted and has an independent reviewer accept the evidence; and what completion still does not mean.
- Voice - the dictation engine ladder and what happens when an engine is unavailable.
- The MCP server - the endpoint file, the token, and how an external process finds a running app; the Vault and the Ledger are what it serves.
- Agent integrations - which coding CLIs get which depth of integration, as a table read from the registries, including what is best-effort.
- Claude Code and Codex together - running both in one workspace: a worktree per pane, mixed swarm roles, per-pane accounts and what the review gate does and does not stop.
- Git worktrees for AI coding agents - why parallel agents need a checkout each, what a worktree does not share, and how the setup is automated.
- Run multiple AI coding agents in parallel - the five steps: split the work, a worktree each, make overlap visible, review, merge one at a time.
- Vibe coding with AI coding agents - the terms (vibe coding, agentic coding, AI pair programming) and a five-step workflow that adds isolation, checks and a record.
- Privacy and telemetry - every network call the app makes and why, what is off by default, and what a crash report contains.
- Troubleshooting - the failures that have actually happened here, and their causes.
How to read these pages
Each page carries the date and commit its claims were read at. That is not decoration: an agentic environment changes faster than its documentation, and a page that cannot be dated cannot be trusted. If a page's stamp is far behind the build you are running, treat the page as a hypothesis and the binary as the answer.
Numbers on these pages are measured on real hardware and say which. Where a step could not be executed on the machine that wrote the page, the page says that instead of quoting a plausible figure.