🌱 OpenSeed v0.1

An open, recursive spec format — not a framework

Write the install once.
Any agent runs it on any machine.

A SEED is a markdown spec for installing your software — its dependencies, the steps to set it up, and read-only checks that prove it worked. Any AI coding agent reads it and installs on the host in front of it, adapting to the environment: apt or brew, the right version pins, the OS’s actual paths. No runtime, no SDK, no lock-in. What the agent shows you is what runs.

Read the spec → View on GitHub
$ plow install git@github.com:you/your-seed.git click to copy · this is what your users run

The format

It’s markdown. An agent reads it top-down.

A fixed grammar: a # Purpose H1, then four required H2s in order, then Feedback if you want install reports. The format is written in itself — an agent that’s read one SEED has read them all. No parser, no runtime to learn.

# Purpose

Purpose

One H1. The what and why, linking to the README pitch.

required
## Dependencies

Dependencies

Other SEEDs and system packages this install needs.

required
## Objects

Objects

The nouns of the system — the things that exist once it’s installed.

required
## Actions

Actions

The steps that set it up. The agent runs these, showing each command first.

required
## Verification

Verification

Read-only checks that prove the install worked.

required
## Feedback

Feedback

Opt-in install reports that flow back to improve the spec.

optional

Purpose and the four required H2s come first, in order. Optional sections follow as Feedback → Open Items → Non-Goals, and a root SEED opens with a ## Normative Language declaration (sub-SEEDs inherit it).

Why publish to the convention

OpenSeed gives you four capabilities.

01

Runs on any agent, any system — no lock-in.

One spec, every host. The agent picks apt or brew, the right pins, the correct paths for the OS it’s on. The format is a convention nobody owns — there’s no vendor between your spec and the machine. Your individual SEED is a repo you own and fork; the format it conforms to is open and ownerless. Two different things — neither one locks you in.

Claude CodeCodexOpenClawHermesGemini CLICursorAgy 
linux ❯ apt-get install ffmpeg ✓
macos ❯ brew install ffmpeg     ✓
one spec → apt or brew, whichever the host has
02

OpenSeeds compose, recursively.

A SEED can depend on other SEEDs. The agent installs leaves-first, so complex deployments are just SEEDs that name other SEEDs — no orchestration layer. Each node is published and verified on its own, and every SEED you ship becomes a dependency someone else can build on.

seed-hostex-history-ingest
└─ seed-hermes-gbrain
   └─ seed-hermes
      └─ (leaves first)
03

Reviewable by construction.

Security in OpenSeed is review, not a sandbox. The convention requires a conforming agent to:

  • Show every command before running itthe spec mandates display-then-execute for all install shell, so a reviewer reads what runs
  • Keep secrets out of the specthe format bans literal keys, tokens, and userinfo, and requires presence-only probes during authoring
  • Audit a fresh clone before executing itclone-mode installs run a pre-execution exfiltration / backdoor checklist and abort on a high-confidence finding

These are obligations the convention places on the agent, enforced by review — not a promise a runtime can’t be bypassed.

04

Verifiable, not just installable.

Every SEED ends with ## Verification — read-only checks that confirm the install actually worked, not just that the steps ran. For CI or any non-agent caller, a SEED can ship a ref/verify.sh that runs a few structural checks and returns a plain exit code.

$ bash ref/verify.sh
tree conforms ✓  exit 0
installed ≠ works. Verification proves both — and anyone can fork the checks.

Compatible tooling

Works with Plow — a package manager built for OpenSeed.

An OpenSeed is just markdown — anyone can author, host, and install one by hand. Plow gives SEEDs a home: a one-line installer, and a registry to discover and fork SEEDs others have published.

One-line install

Curl the installer; Plow drops the OpenSeed skills into every agent on your machine.

Discover & fork

Browse a registry of published SEEDs — find one, fork it, build on it.

Specs that improve

Opt-in, anonymous install reports are coming — they’ll feed real-world install signal back into the spec so authors can tighten it.

$curl -fsSL plow.co/install.sh | bashclick to copy

Install reporting is opt-in and off unless a SEED declares it. The convention works without Plow; Plow just makes it effortless.

Write the install once.
Trust that any agent runs it.

Read the spec on GitHub → Get Plow

A community convention, shaped in the open. Open to all — propose a change.