Skip to content

For developers

The rest of this site is written for someone using palaia, on purpose — this page is the exception, for anyone building against it, extending it, or just curious what’s actually happening under the plain-language explanations elsewhere.

What this guide calls a memory is what the engineering documentation calls a vault: a directory of Markdown files with a small YAML block at the top of each one, notes linking each other by relative path, and the whole thing versioned with git. The full, normative format — every field, every rule for what makes a file valid — is in the repository:

  • v3/docs/ — the format itself is fully specified in the file named vault-format.md there.
  • v3/decisions/ — the recorded decisions behind it, including why it’s clean-room rather than derived from any existing project’s code.

Because it’s just files, you can read, script against, or version-control a memory with ordinary tools — nothing here requires palaia’s own code to be running.

Every AI tool in Connect your AI talks to palaia over MCP (the Model Context Protocol) — an open, cross-vendor standard for how an AI tool discovers and calls tools, not something specific to palaia. Anything that speaks it can connect, including a tool you write yourself; see Any other AI app for the bare connection details a new client needs.

Sign-in for tools that connect from outside your own network follows the standard OAuth 2.1 flow with dynamic client registration — again, not a palaia invention, so any client library that already speaks it works without palaia-specific code.

The marketplace’s tools are built against a small, independently installable SDK — it has no dependency on the hub’s own code, so it stays usable even if you’re not running palaia’s server at all. It covers what a tool declares about itself (what it needs from a user, what it’s allowed to touch) and the submission checks a listing has to pass, including the same plain-language check this whole site is written to satisfy.

Everything that happens in a hub — a note saved, a proposal created, a session connecting — is published as an event a script or an automation can react to, including outbound messages to another system. The event shapes and the safe subset of conditions an automation can express (never a general scripting language — a deliberate limit) are documented at v3/docs/events.md.

The version before this one is still fully supported, on its own separate branch, for anyone not ready to move. If you’re one of its users, start with the migration guide instead — it covers what carries over, what changes, and what this version doesn’t do yet.

The hub itself — the program this whole site is documentation for — is open source. Its own engineering documentation, architecture notes, and the full specification set that built it live at the root of the repository:

  • v3/MASTERPLAN.md — what it is and why it’s built the way it is.
  • v3/docs/ — every subsystem’s own reference.
  • v3/specs/ — the work packages, if you want the full history of a decision rather than just its outcome.