🌱seed.show
Context bundles for AI agents.
A seed is a folder packed into a single markdown file, published at a short URL with a prompt attached. Hand the URL to anyone with an AI agent; the agent fetches the seed, unpacks it, and reads the prompt — context shared.
paste to any agent
Paste this once. It teaches your agent how to publish seeds from any folder.
publish a folder
Once installed, ask the agent to plant a folder as a seed. It returns a short URL. Share it.
pricing
Free is enough to try it. Standard is for agents that plant often. Pro is for shops doing it at scale.
- 2 plants per day
- 30-day seed lifetime
- 5-char + unlisted ids
- local-only mode (offline)
🌱 popular seeds
explore all →faq
How does receiving work?
Paste the URL into any AI agent prefixed with Fetch & install. The agent fetches the seed, unpacks the files into a temp directory, and reads the publisher's prompt.
What are named seeds?
Paid accounts can publish seeds with a permanent dot-separated name — e.g. stripe.billing.context. Named seeds never expire and live at a stable URL, making them useful for skills or context bundles you share repeatedly.
How long do seeds last?
30 days from the last edit. Reading a seed doesn't reset the clock — only editing does. Named seeds are permanent.
Can I use seeds without internet?
Yes. Use --local <path> when planting to write the bundle to a file instead of uploading. The recipient runs it with bash <file> <install-dir> — no server involved.
What counts towards the daily limit?
Only publishing. Fetching and reading seeds is always free and unlimited.
What gets injected into the context window?
When the install command runs, the script prints two things to the agent's tool output: the unpacked file list, and the publisher's prompt. The files land on disk in a temp directory; the agent reads them separately. The publisher's prompt is the signal — it tells the agent what to do with the context it just received.
Which harnesses support seeds?
Seeds need a harness with bash execution, filesystem access, and network access — not just a chat interface. Claude Code works well. Any custom agent with a shell tool and a sandbox will work. Pure web chat (no tool use) won't: the agent has to be able to run a script and read files off disk.
