One calm place for everything you ask AI to do.
An assistant is a Markdown file. A workspace is a folder. A conversation is a real
claude, codex, pi, opencode or
gemini process on your own machine — nothing simulated.
You already own capable agent CLIs. What you do not own is a place to keep them: their instructions, their files, what they cost, which machine they ran on, and what they did yesterday. Entity is that place.
Every assistant is a plain .md file in a folder you can open in any editor. Every
answer is a real process with a real exit code, a transcript you can read and a price
the harness actually reported. Nothing is synced to a service you did not choose.
- Your files, your machine. Workspaces are directories under
~/.entity/; nothing is uploaded unless you connect a server yourself. - The instruction is a file. Edit it, version it, hand it to someone, keep it forever. A frontmatter block decides harness, model, folder, skills and tools.
- Nothing is simulated. Missing binary or a signed-out CLI is reported as a failed turn, not a friendly fake answer.
- Answers you can audit. Every step keeps its thought, tool output, duration and tokens. One click shows where the time and the money went.
- It is not only for code. Campaigns, budgets, itineraries, research memos, support inboxes — the screenshots below are one workspace doing all of it.
- Your own CLIs stay yours. No wrapper, no proxy, no rewriting of their config.
Also in the box: an execution timeline per answer · a links-and-paths drawer
derived from each thread · delegation cards and a session tree behind spawn N /
spawn per · @entity, @thread:, @artifact:, @maka:, @youtrack: and
@orca: mentions resolved before the agent sees them · personal and shared artifact
shelves with a two-clock archive sweep · a folder-per-thread base directory with
/project: skills · a terminal beside the chat (Ctrl+` ) · voice with barge-in and
read-aloud · ZIP backups on a rolling schedule · one-verb agent CLI with an audit log ·
two-way sync with a shared server, evaluator included · a command palette and a
keyboard guide that cannot drift from the handlers.
| you want to… | how it looks here |
|---|---|
| Run a small content studio | one workspace, an entity per role (brand writer, editor, researcher), shared Notion and Slack servers, skills for voice and structure — a launch campaign produced end to end |
| Review numbers every week | an entity with dir: pointing at the finance folder and a budget skill; answers land as artifacts on the shelf with the chart inline |
| Keep research you can cite | @maka: pulls notes from your vault, @youtrack: scopes issues to a project, and the links drawer collects every reference the thread touched |
| Give a coding agent a home | one entity per repo with dir: and its own skill set, @file: mention for files, a terminal in the same folder, spend per repo in the charts |
| Share assistants with a team | connect a workspace to a server, members sign in as themselves, MCP servers and skills travel with the workspace, and the guard scans everything before it is written |
| Keep a shelf without hoarding | saves land personal, one click shares a file with the room, and two clocks archive then delete by last opened — while a live thread keeps the files it still names |
| Fan one instruction across a list | spawn per @youtrack:… @coder fix it turns every matching issue into its own session, card and continuation, with the rows painted as they run |
| Keep the office on your machines | pin a thread to an ssh host and the harness runs there, inventories from one read-only scan, nothing installed on the server |
Works with the assistant CLIs you already have. Entity finds them on PATH, and an
assistant's Configuration dialog can pin a machine-wide executable per harness. They
are installed from a terminal once; after that you never need one to use the app.
Claude Code · claude
Codex CLI · codex
Gemini CLI · gemini
opencode · opencode
pi · pi
A run is a real process with a real exit code. Entity never invents a reply: if the binary is missing or not signed in, the turn says so.
Requirements: Node.js ^20.19 or >=22.12, plus one of the CLIs above.
git clone https://github.com/agung2001/entity.git
cd entity
npm install
npm run dev:desktop # Vite + Electron, HMR includedOn a fresh profile the app lands on a three-step welcome card: name the room,
6875
pick
an agent CLI, say hello. It creates the machine-local owner on the way in, so there is
no sign-in form to fill. Prefer a scratch profile?
ENTITY_HOME=/tmp/entity-scratch npm run dev:desktop.
| Command | What it does |
|---|---|
npm run dev |
Renderer only, in a browser at http://localhost:5183 |
npm run dev:lan |
The same, bound to 0.0.0.0 so another machine on your network can open it |
npm run dev:desktop |
Vite and Electron, pointed at the dev server |
npm run test |
Unit tests + the static UI contract |
npm run test:gate |
Unit + smoke — the release gate |
npm run package |
A distributable for the current platform → release/ |
Full command list, ports, test packs and release builds: Quickstart →
| Document | What is in it |
|---|---|
| Quickstart | requirements, every npm command, ports, LAN dev, profiles, test packs, tagged releases |
| Onboarding | the first run: three steps, the silent owner account, and the return launch |
| Workspaces and entities | the folder tree, the entity file, Learning, chat retention, moving in from another tool |
| Conversations | transcripts, timeline, mentions, delegation sessions and fan-out, base directories, usage and activity |
| Skills | how /project:, /workspace: and /<harness>: resolve, attach versus link |
| MCP servers | one set per workspace, the server page and its tool list, how each harness receives it, the agent-facing endpoint |
| Artifacts and backups | personal versus shared shelves, share and unshare, the archive sweep, artifact://, ZIP snapshots |
| Machines | ~/.ssh/config, test & scan, the machine that sweeps, pinned threads, the terminal |
| Voice | barge-in, transcription engines, privacy, sound cues |
| Remote workspaces | two-way sync, the three-way diff, the guard and its evaluator, the artifact scope index |
| Agent CLI | the verbs an entity can call on its own workspace |
| Deploy the server | Docker, GHCR image, reverse proxy, hardening |
| Architecture | layers, the one call table, child processes, the server |
| Security | trust boundaries, what a run may write, credentials |
| Limits | what it deliberately does not do |
| Status | what "growing" means here, and what is out of scope |
| Migrating from Multica | a real 8-agent port, step by step |
| README media | how the demos and the animated mark are recorded, encoded and checked |
DESIGN.md |
the interface contract every UI change is held to |
Built with Vue 3, Vite and Electron. Credits · Status · Limits · How these demos are made