About the protocol.
The Factlet Protocol is an open specification, governed by a working group, with a reference implementation maintained by Kernora. This page explains who maintains what, how decisions are made, and how to get involved.
Maintainers
v0.1 of the specification was authored by Mihir Choudhary, founder of Kernora. The reference Python SDK is maintained by Kernora as a community service. The spec itself is owned by the protocol working group; future versions reflect community RFCs, not vendor decisions.
How decisions are made
Substantive changes to the spec — new primitives, breaking schema changes, field renames — flow through the open RFC process at github.com/factlet-ai/spec/discussions. Each accepted RFC includes the rationale, alternatives considered, and the migration path if breaking. Discussion threads stay open even after a decision lands; the audit trail is the spec's most durable artifact.
Reference SDK
Kernora maintains the canonical Python reference SDK at github.com/factlet-ai/reference-sdk. The reference is what implementations test against; it stays intentionally minimal and dependency-light (~400 lines, 15 passing tests, only PyYAML required). TypeScript reference SDK is planned for v0.1.1 after Python stabilizes.
About Nora — the production reference implementation
Nora is the production reference implementation of the Factlet Protocol, built and maintained by Kernora. It runs alongside your AI coding tools — captures every session, extracts architectural decisions, and feeds the resulting Factbook back into the next session.
Where the Python reference SDK is the smallest possible thing that demonstrates the protocol contract, Nora is the production version: embedding-based retrieval, a 19-tool MCP server, IDE extensions for Claude Code / Cursor / Kiro / VS Code, a macOS desktop app, real-time FactSignal scoring, and the low-FactSignal warning surfaced as an in-IDE banner.
Nora is open source and local-first. No cloud required. No API key required by default. Bring your own model — Claude, GPT, Gemini, or a local model on your machine.
- Source code
- Project home
- Other tools welcome
Nora is one implementation among many that the protocol allows. Cursor, Claude Code, Continue.dev, Aider, Goose, OpenCode — any tool with retrieval can implement a reader in an afternoon and read the same Factbooks.
Get involved
- Implement a reader
If you build a tool that does retrieval — IDE plugin, internal copilot, agent framework — the SDK is ~400 lines and a reader is ~80 of them. github.com/factlet-ai/reference-sdk
- Open an RFC
Push back on the spec via spec discussions. v0.2 ships within 90 days. Field names, scoring algorithms, supersession semantics — every decision is up for revision until v1.0 locks.
- Contribute a Factbook
Add a sanitized example to the registry. Any real-world domain helps implementers calibrate. See the existing payments / frontend / ML pipeline examples to crib from.
Contact
Spec questions and RFCs: github.com/factlet-ai/spec/discussions. Press, partnership, security disclosures, or anything else: hello@kernora.ai.