Nervous system
How it knows your code. 3 ways to find an answer: language server, parsing, then text search.
The Genome
A live, ranked map of your files and their dependencies, a slice of which goes into every prompt.
Language servers
The same servers your editor uses, installed in the background for the languages in your project.
Code intelligence
Questions like “who calls parseConfig?” are answered by symbol, with parsing and text search as fallbacks.
What changes together
Files that usually change together in git are pulled in together, and recent edits get a boost.
The MarionetteExperimental
An optional pass where a cheap model finds the files and lines your request is about.
Hands
How it changes code. 8 compound tools that each finish a whole job in one call.
Edits by name
TypeScript and JavaScript are changed by symbol, so edits cannot fail on whitespace or quotes.
Renames and moves
Rename symbols everywhere, move code and files without breaking imports, extract functions.
Compound tools
One call finishes a whole job, like a rename that confirms nothing still points at the old name.
Scripted changes
For bigger jobs it writes one script that edits code in any language, checks the result and reports back.
Plan mode
For changes across many files it writes a plan first and edits only once you approve.
Neovim inside
A real Neovim in the terminal app, with your config, plugins and language servers.
Python kernel
A Python interpreter that keeps its data and imports between steps. Off by default.
Immune system
How it catches mistakes. 405 automated checks try to break Empryo itself.
Checks after every edit
Errors are counted before and after each edit, so the agent hears what it introduced and what it fixed.
The project tool
Runs the right lint, typecheck, test, build and format commands across 23 ecosystems.
The reviewer
A separate agent with a clean context reads what changed and gives one verdict, on demand or as a CI gate.
The goal loop
Say once what done looks like; it works in rounds and a fresh reviewer judges each one.
Immunity
405 automated checks try to break Empryo, 298 of them on all three platforms.
Memory
What it keeps between sessions. 3 memories at most join a turn, and only when they help.
Project memory
Preferences, decisions and gotchas, kept for one project or all of them, saved on its own when it matters.
Recall that fits the turn
Each turn adds up to three memories that match your message and files. Nothing relevant, nothing added.
Cleanup
The memory browser flags duplicates, memories about deleted files and ones that rarely help.
Project rules
Standing rules live in EMPRYO.md; instruction files written for Claude Code, Cursor and others are read too.
Senses
How it looks beyond your files. Local addresses only, by default, for the built-in browser.
Web search
Searches the web for documentation, APIs and error messages, and fetches readable pages.
The browser
Opens the app you are building, clicks through it and takes screenshots to check a change works.
Computer use
Operates its own browser, your Chrome, another app or the whole screen, with permissions and a stop button.
Provider tools
Code execution, provider-side search and Claude's own editor, wherever the model supports them.
Voice and reach
How you talk to it, from anywhere. 3 surfaces, one agent: terminal, desktop and headless.
Voice input
Speak a prompt, or go hands-free so each pause sends. Transcribe on your ChatGPT plan, watch words appear live with an OpenAI key, or stay offline with whisper.cpp.
Steering
Type while it works. Your message lands at its next step, so you never have to start over.
HearthDeprecated
Sent tasks and approved edits from Telegram or Discord. It is switched off and only prints a notice now.
Three surfaces
A terminal app, a desktop app, or no UI at all, all running the same agent and settings.
WSL repos
Open a repo that lives in WSL and Empryo runs inside the distro, so git, language servers and every command use its Linux tools.
Headless and CI
Pipe in a prompt, get text, JSON or live events back, and fail a pipeline on a review verdict.
Five languages
Use Empryo in English, Arabic, French, Korean or Russian, with a mirrored desktop layout for Arabic.
Metabolism
How it spends your money. /login one picker for the plans you already pay for.
Subscriptions
Sign in to Claude, ChatGPT, Copilot, Cursor, Devin, Gemini, Grok, Kimi and more with no add-on.
Keys and local models
Or save an API key, or run a local model. Many providers are supported.
The task router
A cheap model searches and summarizes while a strong one writes the code.
Cost tracking
What every session costs as you work, by model and helper, with live meters for subscription plans.
Fallback chains
Each model gets its own backups, so an overloaded provider does not stop your session.
Thinking and effort
Choose how much the model thinks before answering, per model or provider, and tune retries.
Growth
How it becomes yours. 1 JSON file is all a new panel takes.
Cells and MorphsExperimental
One JSON file adds a panel, status chip, command or shortcut to the desktop app.
Skills
A SKILL.md file teaches the agent a framework, workflow or style guide, turned on per session.
MCP servers
Connect servers over stdio, HTTP or SSE, and the agent uses their tools next to its own.
Hooks
Run your own commands on tool use or session start and end. Claude Code hooks work unchanged.
Your own agents
Define specialist subagents as markdown files. Claude Code, opencode and Gemini CLI agents are read too.
Themes
Over 40 built-in colour themes, or your own JSON theme that reloads as you edit it.
Presets
Shareable settings bundles by name, file or URL, so setups match across machines and teams.
Colonies
Many agents at once. 32 tabs in a desktop window, 10 in the terminal, each running its own agent.
Tabs
Agents side by side in one project, warning each other before editing the same file.
Worktrees
A tab can work in its own git checkout and branch, with its own engine and Genome. It moves between checkouts with its chat, and the agent can open one for a task.
Parallel helpers
The main agent splits big tasks across helpers that run in parallel and can use cheaper models.
Councils
Two to four agents with different briefs debate one question and reach a verdict.
Agent graphs
A graph of agents with loops, review steps and pauses for your answer.
Duck and advisor
A duck to think out loud with, and an advisor for a second opinion. Neither is the coding agent.
Docs for other agents
Copy a page, connect the read-only docs MCP server, or use the empryo skill that ships with the app.
Time
How it goes back, and keeps going. 76× smaller, at most, when a long session folds into a summary.
Checkpoints
A checkpoint after every turn. Step back, undo, or start again from any earlier point.
Sessions
Every conversation is saved as you go. Resume it, export it, or move it to another machine.
Compaction
Near the end of the model's window, older turns fold into a summary in about a second.
Routines
Schedule runs with /schedule: a nightly review, an hourly check. They run while Empryo is open, and you can stop a run at any time.
Meet the whole creature
Everything on this page ships in one app for macOS, Windows and Linux.