Real terminals, not prompts
Every employee is a live Claude Code, Codex, OpenCode, Gemini CLI or Aider session in its own sandboxed folder. Open several side by side, type into them, interrupt them.
Give one goal. The CEO splits it into tasks, department leads do the work in real coding terminals, QA reviews the diff, rules are enforced by code, and an evidence-backed report lands in an Obsidian vault you can open yourself.
Works with the tools and backends you already use
How it works
Claude Code, Codex, OpenCode, Gemini CLI or Aider on PATH. Log them into your subscription or add API keys in Settings → Backends.
A template staffs management, software, design, research, QA, marketing, client relations, finance, operations and administration — or start empty and let the CEO hire.
The secretary carries it to the CEO. The CEO writes a plan with acceptance criteria per task; in Co-Founder mode it waits for your approval.
Leads work in a shared project repository. Click anyone to look over their shoulder, message them, or read what they wrote to each other.
QA reviews the diff, the rule engine checks paths, dependencies and commands, and a report with real evidence is filed in the vault and committed to git.
Everyone remembers: daily notes, lessons, decisions. After a restart the terminals you opened yourself come back with the tool's continue flag, and orchestrated work is re-dispatched with its brief.
Orchestration
The general manager hands out every job by how hard it is. Each employee works with the right model, at the right effort, in its own files - nobody sings over anyone else.
Claude Code employees can open subagents to go faster, but only the company's three: explorer (read only), implementer and verifier. Company policy sets their model and effort, never above the employee's own model.
The plan stamps every job easy, medium or hard. Easy work opens on Sonnet at low effort, hard work on Opus at high effort: simple jobs don't burn the expensive model, hard ones don't stall on a weak one.
The general manager, department leads and reviewers run on the strongest model by default. Moving one to a weaker model asks for your approval.
When the jobs are done, QA tests the combined result; every bug becomes its own fix job and nobody reviews their own work. On big jobs a refuter subagent challenges each finding, and only the ones that stand reach you.
The 5-hour and weekly windows of the Claude subscription are watched separately. As one fills the company slows down, at the limit it stops, and it carries on by itself when the window resets.
A quiet employee gets a status check from its lead, and a stuck session reopens where it left off. Unsaved work is snapshotted every 10 minutes, and git commands that would wipe someone else's work are blocked.
Employees announce the files they work on, and parallel subagents never share files. An unknown file is not deleted; its owner is asked first, and you when nobody owns it.
Messages travel through a background queue; when an employee's terminal is closed, its messages are delivered as soon as it opens. Claude employees message each other through Claude's own messaging, and it all shows up on the Messages page.
The platform
One goal in — an org that plans itself, works in the open, remembers, and holds its own budget.
Every employee is a live Claude Code, Codex, OpenCode, Gemini CLI or Aider session in its own sandboxed folder. Open several side by side, type into them, interrupt them.
The CEO plans, hands tasks to department leads, hires missing specialists and opens departments when the mission needs them.
Profiles, daily logs, lessons, decisions and reports are Markdown notes with wikilinks. You can read and edit them; retrieval costs zero tokens.
Your decisions become policies checked against the real diff and the commands that ran. Violations send work back with the violation listed.
Reports carry the diff, raw command output, the rationale (why this path, what was rejected), review verdict and cost.
Fully autonomous, or human in the loop: plans, deliveries and hires wait in your approval queue with a note back to the employee.
Direct conversations and department channels. Orchestration is narrated as messages; agents can message each other and you.
The local gateway measures every call per employee from the provider's own counters. Monthly and per-person caps stop dispatch, not running work.
GitHub, Slack, Notion, Linear, Supabase, Stripe, Sentry, Playwright and more via MCP — one catalog, secrets kept in the key store, written into every tool's own format.
Walking, speech bubbles and status come from real events. Rooms and buildings are editable. Positions are simulated on the server, so every screen sees the same office.
Several keys per backend; on 429 the gateway moves to the next, on 401 it retires the key. Tools never see real keys.
LAN access with a PIN, one-click Cloudflare tunnel, single-pane mobile layout. Turkish and English.
0online right now
0terminals running
45downloads
Why this one
Most multi-agent tools are a black box: eight agents run, one gets stuck, nothing is remembered. Agentic Company makes the company observable, accountable and persistent.
| Agentic Company | Typical agent frameworks | |
|---|---|---|
| Agents | Real CLI terminals you can open and type into | API loops behind a chat window |
| Memory | Obsidian vault, human-readable, git-friendly | Vector store or nothing |
| Rules | Deterministic checks on the real diff and commands | A line in the system prompt |
| Reports | Diff, raw output, rationale, review, cost | The model's summary of itself |
| Cost | Measured per employee at the gateway; caps | Whatever the bill says |
| Keys | Pool with failover, encrypted at rest, never shown to tools | One key in an env var |
| Deployment | One binary, no runtime, everything on your machine | Node/Python stack, cloud login |
Pricing
The app is a single binary that runs on your machine; you pay your AI provider directly. Paid plans unlock scale, remote access and team features. No usage-based fees from us, ever.
For now plans are bought in the mobile app only. Want one right away? Write to us.
For now plans are bought in the mobile app only. Want one right away? Write to us.
Prices are per install, billed monthly, cancel anytime. Local-first stays true on every plan: nothing about your company leaves your machine, including on Team. Model usage is billed by your provider, not by us. Refund Policy · Terms of Service · Privacy Policy
Download
Single binary. No installer, no runtime. Data stays in your user profile.
x64 · arm64
Run agentic-company.exe; it prints the address and a PIN and opens a window.
Apple Silicon · Intel
Unzip and open Agentic Company.app, or run the binary.
x64 · arm64, statically linked
install-linux.sh adds a desktop entry.
Your company in your pocket: remote access through the tunnel, notifications, team and plan management. Works alongside the desktop app.
agentic-company --port 7777 # LAN access with a PIN agentic-company --tls-cert c.pem --tls-key k.pem agentic-company --backup C:\backups\ac.zip
No. Everything lives under your data directory: company config, the Obsidian vault, a SQLite database, project repositories and workspaces. The only network traffic is your coding tools calling the AI provider you connect (through a local gateway that injects the keys) and signing in with your account. If you use dictation, your voice is turned into text on Chrome's or Edge's own servers.
Tools: Claude Code, Codex CLI, OpenCode, Gemini CLI, Abacus.AI CLI, Aider. Backends: your subscription login, Anthropic API, Google Gemini API, OpenRouter, NVIDIA NIM, Kimi, GLM, MiniMax, any OpenAI-compatible endpoint, or a local Ollama. Tool and backend are chosen per employee.
A goal becomes a plan task for the CEO. The plan lists tasks with department, brief and acceptance criteria — and, when the current team cannot cover the work, new hires. Tasks go to department leads, deliveries go to a reviewer, reports are filed. In Co-Founder mode the plan, every delivery and every hire wait for your approval.
The app is free. You pay your AI provider. Employees on a subscription login cost nothing extra; API-backed employees are measured per call at the gateway, with monthly and per-person caps.
Yes. Open the printed LAN address and enter the PIN, or start the Cloudflare tunnel from Settings for a temporary https address. The layout switches to single-pane tabs on narrow screens.
PIN login with rate limiting, same-origin checks on every write, strict Content-Security-Policy, loopback-only gateway, secrets encrypted at rest on Windows and redacted from reports. The operations runbook that ships with the app documents the full model.