How do I share an agent?

Send the file. Text it, email it, put it in a shared folder, or burn it to a CD like it's 2005. No accounts, no signups, and no one's permission; the recipient runs it with their own model and API keys.

Agents, localized.

An AI agent that belongs to you. Run it on any model, on any computer, or send it in an email.

Get started

Fig. 1. Where an agent lives. At t = 0 its parts are spread across systems run by other people. At t = 2026 they are held in one file, ava.adf.

(i)

The localized state is drawn as a hydrogen 3d_z² orbital, seen from about 35° above its equator so the ring of the orbital reads around its waist. The shading is |ψ|² integrated along the line of sight, and the outline is the contour of that projected density which encloses 90% of the probability. The dots are samples from it, and the two tones mark the sign of Re ψ. The curve is the density of the dots projected onto x. The small one that follows you down the page is the same drawing at a smaller scale.

1Who controls your agent

AI agents read your email, manage your calendar and act on your behalf → they are extensions of you.

Most run inside someone else's product, where that company decides what they remember and which instructions they follow.

With ADF the agent is a file you own. Run it on your laptop, a home server, or the cloud.

2Localization

An agent's memory, instructions, identity keys and history live in a single SQLite file. Copy the file and you have copied the agent.

agent = memory + instructions + identity + history(1)

ADF spec §1.3, One file, one agent
adf_files: mind.md, soul.md

Fig. 2Parts of an agent gathered into one file

3Defined schema

The spec says exactly what goes in the file, so anyone can build software that opens your agent.

ADF spec §3.2, Protected schema
12 adf_ tables + local_*

Fig. 3One file, opened by Studio, the CLI or your own software

4Model independence

Point the agent at another provider or a local model and it keeps its memory, tools and skills.

ADF spec §16, Portability
needs native tool calling

Fig. 4One file, any tool-calling model

5Example agents

Table 1. Four agents, each a single file.

File	What it does
inbox.adf	Sorts your email and calendar and messages you on Telegram when something needs you.
research.adf	Watches the sources you pick and sends a briefing on a schedule.
coach.adf	Builds a training plan, logs every session and adjusts from the log.
ops.adf	Triages incoming work for a small business and asks before it acts.

Send the file to someone and they open it with their own model and API keys.

6For developers

Open-source agent frameworks still keep the agent's state in whatever database the host runs. ADF standardizes the state itself, so the agent outlives the runtime.

6.1Get started

ADF StudioDesktop app for Windows, macOS and Linux: Version 0.7.529 September 2026Release notes
DownloadsWindows.exe macOS (Apple Silicon).dmg macOS (Universal).dmg Linux (Debian, Ubuntu).deb Linux (AppImage).AppImage
ADF CLITerminal app and daemon: npm install -g @agentdocumentformat/cli
SpecificationThe file format: ADF v0.2Draft
APIHTTP API of the ADF daemon: Daemon APIOpenAPI 3

7FAQ

Is ADF a standard?

It's an open specification, currently a v0.2 draft. It's MIT licensed and will change before it settles.

Do I need Studio to use an .adf file?

No. An .adf file is a SQLite database and the spec is public; ADF Studio and the ADF CLI are the two runtimes today.

Which models work?

Any model with native tool calling, hosted or local. Switching models keeps the memory, tools and history, since those live in the file.

Is ava Clippy?

No. Clippy lived inside Word and couldn't leave; ava is its own file and can go anywhere you take it.

We're building agents that belong to the people they work for.

The spec is a v0.2 draft and will change, but ADF Studio and the CLI run agents today. It's all MIT licensed.

Use it, build on it, or tell us what's wrong on GitHub.

Set in STIX Two Text and Inter Tight. Figures drawn on a 24 px grid.