← All products
The project's memory, out of the repo — prompts with a history, and a knowledge site your agent keeps.
A codebase accumulates two kinds of amnesia. Its prompts — the sentences it sends to a model in production — live as string literals with no history, edited in place, reviewed by nobody. And its knowledge — what this thing is, how it runs, why it's built this way — is scattered across Markdown files that were true the week they were written and loaded wholesale into every agent session since.
systemprompt moves both out of the repo and into a versioned store the agent reads on demand and writes as it works.
Every prompt is named, every edit is a new version, and the current one is fetched at runtime — from code with the SDK, or by the agent over MCP. Roll back in one call. Usage and cost telemetry come along for free.
sp.getPrompt("hello")Each project gets a live site of self-contained HTML pages —
overview, architecture, decisions, changelog — that the agent
seeds by studying the repo, then updates as a side effect of the
work. Humans read it at a URL. Agents read it instead of a pile
of stale .md files.
Run the installer in any repo. It registers the MCP server, keeps the key out of version control, and leaves a skill that changes how the agent behaves from then on: read the site for context before working, update the affected pages after meaningful changes, append to the changelog, and hand the human a URL instead of a wall of text.
.mcp.json — the systemprompt MCP server, eight toolsThe dashboard is where a human signs up, mints a key, and edits by hand when they want to. The rest of the time the site simply stays true, because keeping it true is part of the agent's job.