agentlas-os

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

---

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.

---

Questions you probably have

Every place that sells agents ranks them. That is how they make money. People click what is near the top, and the company decides what is near the top.

We don't build a ranking. The schema pins allowHistoryEvidence to a constant false, and a source that returns a scored list has its whole answer refused.

Ratings, install counts, last month's sales — we read none of them. You get the candidates as they are, and your model picks.

No. Alongside backend, security and QA, the role ontology carries an insurance actuary, an M&A diligence lead, an underwriting diligence specialist and a travel planner.

Insurance alone has four communities. Domains outside the list are not blocked.

---

We added our own idea, then removed it

We tried the publisher's own "call me when…" sentences as a ranking signal. The share of searches that put the right agent first fell from 73.1% to 57.0%. We removed it. Those sentences now reserve three seats at the end of the list — the right not to be deleted, never the authority to re-rank.

Asked in Korean, the right agent came back 144th. The same question in English put it 1st. So the order is written in English, and the language of the deliverable is a separate field.

Showing 10, 20, 30 and 50 candidates put the right one on the list 73, 83, 87 and 93% of the time. The default is 30.

---

Desktop is the host itself, so it decides in code — before a run starts — whether the work needs one agent or a team. That decision is recorded as a reserved run, not a suggestion. It staffs only the agents you already installed, and bringing in paid outside help always asks first.

Every LLM you use, in one place: the coding CLIs you already subscribe to (Claude Code, Codex, Antigravity/Gemini, Cursor, Grok, Kimi), your own API keys across 11 providers, and local models. A verified llama.cpp engine ships with it, pulls GGUF models from Hugging Face, and it also talks to LM Studio and MLX.

And when the model you are on runs out of quota, the turn is re-run on the next model in your own priority order. Your saved model choice is not touched, and the original model comes back on its own once the limit clears.

Surface	Role
Agentlas Desktop	Visual local OS for agent teams, memory, browser work and Hub specialists
Hephaestus plugin	This repo — the open-source engine and command surface for your CLI hosts
Agentlas Hub	Public surface for publishing and borrowing specialists
Agentlas Cloud	Owner-scoped store for privately saving and retrieving your own agents

---

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

---

Install: every method, and what it writes

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

One line, no sudo, nothing written outside $HOME. Then restart your agent host and run hephaestus doctor — it prints what is wired and what is not. Prefer a GUI? Agentlas Desktop runs the same runtime on macOS, Windows and Linux.

It configures every supported host it finds on this machine, not only the one you are typing in — detection is by what is on your PATH.

Path	What lands there
~/.agentlas/	the runtime itself, and its generations
~/.local/bin/	the hephaestus and hep-* shell wrappers
~/.claude/, ~/.codex/	plugin registration, commands/prompts, memory hooks
~/.cursor/mcp.json	MCP server registration
~/.gemini/	extension + MCP config
~/.openclaw/, ~/.grok/	skills and hooks
~/.gemini/antigravity/, ~/.gemini/config/	Antigravity workflows and plugin/MCP config — Antigravity lives under .gemini, not .antigravity
~/.config/opencode/, ~/.config/goose/	commands, plugin, MCP config
~/.copilot/mcp-config.json	MCP server registration
~/.agents/	the vendor-neutral skills + plugin directory

Two things that are easy to assume wrongly:

• **It configures every supported host it finds, not just the one you are