New: the RAC specification
A strict semantic layer over OKF.
Google Cloud's Open Knowledge Format standardizes the carrier for organizational knowledge — deliberately permissive, on purpose. RAC specifies what one high-stakes category of knowledge needs that carrier to guarantee: identity that survives refactoring, a closed status lifecycle, typed relationships with referential integrity, and write-time enforcement.
RAC is the engine underneath: Lore keeps a team's recorded requirements, decisions, and designs as typed Markdown in the repo, validates them in CI, and serves them read-only to coding agents over MCP.
pip install rac-core
How it works
Record
Requirements, decisions, and designs live as typed Markdown in your repo, versioned next to the code.
Validate
CI rejects malformed artifacts, broken links, and references to superseded decisions before they land.
Serve
Your agent queries Lore over MCP — four read-only tools — and cites decisions by ID instead of violating them.
Ask your agent: "Should I add a hard delete to the user model?" — it calls Lore, finds your soft-delete decision, cites it by ID, and proposes the compliant change instead of reintroducing what you removed on purpose.
Serving happens over the lore MCP server; the
rac CLI covers validation, inspection, and ingestion.
Why it's different
Deterministic, not probabilistic
Retrieval makes no model calls and uses no embeddings. The same query returns the same answer, every run.
Read-only by design
Agents cite decisions by ID; they can never mutate the store. Changes land only through human-reviewed pull requests.
Plain Markdown in your repo
No database, no vendor lock-in. Your knowledge is versioned next to your code and readable without any tool.
Enforced in CI, air-gap friendly
Broken links and superseded references are rejected before they land. Runs fully offline — nothing leaves your machine.
The ecosystem
Lore (rac-core)
The engine: the rac CLI, validation gates, and the read-only lore MCP server. This is what you install.
Proofkeeper
Turns each recorded capability into a re-runnable Playwright test, proposed back by pull request.
Wayfinder
Deterministic prompt-complexity routing — a hard-or-easy call on every prompt, offline, no model call.
rac-sdk
Non-Python language SDKs — thin clients over the engine's stable
--json contracts.
rac-editors
IDE and editor integrations, one subdir per client.
rac-ci
The CI delivery surface — validation and gating wrappers, GitHub first.
rac-benchmarks
Evaluation suites, one subdir per benchmark.
Plus rac-connectors and everything else — see the full repository map or the itsthelore org on GitHub.
Lore is built on RAC (requirements-as-code), the
open-source engine — pip install rac-core installs the CLI and
the lore MCP server together.
Up and running in five minutes.
Install the CLI, scaffold your first artifact, connect your agent.
Lore is open source and under active development.