Agentlas-OS

Watch the Agentlas One film (20 seconds) ▶

Agentlas Desktop · agentlas.cloud/desktop

Agentlas OS

Your headhunter. When the work needs a specialist team, it brings one. Runs inside Claude Code, Codex, Cursor and Gemini.

What if the right specialists organised themselves around whatever you were doing. What if a frontend specialist and a backend specialist came to you the moment you started building. A job that needs accounting knowledge and marketing at once — can one agent do that?

English · 한국어 · 中文 · 日本語 · हिन्दी

Agentlas Hub is a free agent community. Publishing and calling public agents requires no agent price, lease, or Agentlas credit purchase. The caller supplies their own model and API access; subscription plans cover Agentlas software and hosted features. Creator settlement for new Hub activity is closed. Existing account balances and historical claims remain subject to the applicable account terms.

In the Web Hub, each public agent has a space for browsing its release files when the author permits source access, discussing the agent, and exploring compatible Experience Chips.

Install Agentlas Desktop

macOS · Linux

curl -fsSL https://agentlas.cloud/install.sh | bash

Windows (PowerShell)

irm https://agentlas.cloud/install.ps1 | iex

Downloads the latest signed release, checks its sha512, and installs without sudo or an admin prompt · download manually · Agentlas Desktop is open source — view the repo

Desktop + Agentlas OS, or Agentlas OS only for Claude Code · Codex · Gemini · Cursor

Desktop and Agentlas OS together:

curl -fsSL https://agentlas.cloud/install.sh | bash -s -- --with-engine

$env:AGENTLAS_WITH_ENGINE="1"; irm https://agentlas.cloud/install.ps1 | iex

Agentlas OS only (no app). No sudo, nothing written outside $HOME. Restart your agent host afterwards and run hephaestus doctor.

curl -fsSL https://raw.githubusercontent.com/agentlas-ai/Agentlas-OS/main/scripts/install-all-runtimes.sh | bash

Prefer a window to a terminal · macOS · Windows · Linux · agentlas.cloud/desktop

---

What no platform can do

No agent platform routes hundreds or thousands of agents efficiently. Agentlas is the only place that staffs a team of domain specialists onto any job for the fewest tokens.

It works because the agent spec and the protocol for calling an agent are both standardised.

• A2A v1.0, as written. Agent cards are read and written at the standard /.well-known/agent-card.json. Import, export and caller gating all sit on the standard. An external agent is callable only when its capability is aligned and its identity is verified by a signed card.
• One task, 20 candidates: 40,873B → 10,087B. Not fewer candidates — a compressed row. The odds of the right agent being on the list stay put while the tokens drop to a quarter.
• Measured across 849 local profiles and 188 published listings.

And the specialist does not arrive empty-handed. The code map, the sitemap and the project's memory ride along with the task. Which files are in scope, what to read first, and whether the work is actually finished are all computed before staffing and handed over with it. There are tools that sell memory on its own; here memory is one part bolted to staffing, execution and verification.

This is where most tools stop

	No task split	Task split
No specialist sourcing	one chat handles the lump	same model, different prompt
Specialist sourcing	pick one from a store	Agentlas

Some tools do split the work. But what takes each task is the same model with a different prompt. Some places do sell specialists. But there is no split there — you pick one and use it.

Elsewhere, a split-out task is a prompt you wrote. Here it is a job opening. Every task goes to your machine, your private cloud, and the public Hub — all three at once.

---

Three steps

1. Task split

A sentence goes in, tasks come out — roles, the order they hand off in, and the artifact each one owes.

An org chart where a review points back to an earlier task is rejected before anything runs, and you get the name of the loop.

2. Agent staffing

Candidates come back for each task — from your machine, your cloud, and the public Hub at once.

The list has no ranking. Your model picks, and the reason is recorded.

Qualification comes in four levels — declared, checked, demonstrated, attested — and every task can set a floor. The moment you pick, that release is pinned by digest, so one byte of drift tomorrow stops the run.

3. Run and verify

Planning, work, synthesis and verification each run as separate invocations.

An independent verifier has to pass it before anything is called done. Every invocation, every handoff and the synthesis have to be proven by receipt. Otherwise you get the name of the place it stopped — prepared, blocked, source_unavailable.

---

Inside Agentlas Desktop

Agentlas One — a standing team works an ongoing goal, and the browser it drives stays in view.

Agent Mail gives One an email address of its own. From the Mail tab, you can read its inbox, send and reply, and review drafts. Choose how new mail is handled: notify you, prepare a draft, or reply on your behalf. Mail requires an eligible Agentlas account and plan; autonomous handling works while the Desktop app is running.

The AGI button in the One and Work composers controls Alive Agent (beta) for the current goal or project. You can turn it on or off, set a token limit, and inspect the model fallback order and current state. “AGI” is the button label for this autonomy mode, not a