Give your coding agent the decisions your team already made — so it stops re-doing things you ruled out.
Agents that know why.
pip install requirements-as-code
Requires Python 3.11+ · uv tool install requirements-as-code also works.
What it is
Your agent reintroduces an approach you rejected months ago. It rebuilds something you deliberately removed. The decision was written down — in an ADR nobody, human or agent, ever reopened.
Lore stores your requirements, decisions, designs, and roadmaps as typed Markdown in your repo, and serves them to Claude Code, Cursor, and Claude Desktop over MCP. The agent cites your decisions instead of violating them.
No AI in the core. No inference. No guessing. Just your team's recorded knowledge, in your Git, handed to the agent that needs it.
Lore is built on RAC — Requirements as Code — the open-source engine underneath. For now the package, CLI, and MCP server all ship under the rac name.
Point your agent at your repo and ask:
"Should I add a hard delete to the user model?"
The agent calls Lore, finds your soft-delete decision, cites it by ID, and proposes the compliant change — instead of reintroducing the thing you removed on purpose.
The server exposes four read-only tools: get_artifact, search_artifacts, get_related, get_summary. It never writes to your repo.
▶ Full walkthrough + runnable example: examples/guide/
Why this works
The code is structured, the tests are automated, the infrastructure is versioned — but the reasoning behind what you build is scattered across tickets, chats, and dead docs. Agents can't act on what they can't read, so they re-litigate settled decisions.
Lore puts that reasoning back in the repo as typed, connected artifacts, then serves it to the agent through a deterministic interface. You write the decision once, in Markdown; RAC validates it, links it, and makes it retrievable — durable context for both humans and AI, with no proprietary format and no hosted platform.
How this relates to spec-driven development
Spec-driven development (SDD) tools — GitHub Spec Kit, OpenSpec, Kiro — manage the change: proposal, design, tasks, carried through to implementation. They treat requirements as ephemeral inputs that are consumed and archived. RAC manages the requirements: a durable, versioned, governed corpus that persists across changes and is served to your agent over MCP. RAC is the layer above SDD tools, not a competitor to them — an SDD tool drives each change, while Lore holds the decisions and requirements those changes draw on.
| Dimension | Lore / RAC | GitHub Spec Kit | OpenSpec |
|---|---|---|---|
| Requirement persistence | Requirements, decisions, designs, and roadmaps are long-lived artifacts that persist across changes | Spec, plan, and tasks are created per feature under specs/<feature>/ |
Change folders are archived on completion under openspec/changes/archive/; the specs directory is updated |
| Change management | None — RAC does not manage the change cycle; pair it with an SDD tool | Slash-command workflow: specify, clarify, plan, tasks, implement | Slash-command workflow: propose, apply, archive |
| Traceability | Typed Related links between artifacts; rac relationships --validate checks them in CI |
/speckit.analyze runs cross-artifact consistency and coverage analysis |
openspec validate checks changes and specs for structural issues |
| Tool coupling | Read-only MCP server; works with any MCP client | Slash commands or skills installed per agent at init (GitHub Copilot, Claude Code, Cursor, and others) | Slash commands for 20+ AI assistants |
| Install footprint | pip install requirements-as-code (Python 3.11+) |
uv tool install specify-cli from the Git repository (Python 3.11+) |
npm install -g @fission-ai/openspec (Node.js 20.19+) |
How this relates to OKF
Google's Open Knowledge Format (OKF) standardises the carrier — a Git tree of Markdown with YAML front matter — and is deliberately permissive: consumers must not reject a bundle for missing fields, unknown types, or broken links. OKF says "if you can cat it, you can read it." RAC says "and CI guarantees the file is well-formed, the decision is consistent, and nothing points at a superseded artifact." OKF is read-time interchange; RAC is write-time enforcement — deterministic, cross-artifact validation (referential integrity, status-consistency, illegal-edge detection) that fails your build before bad knowledge lands. A RAC repo is a conformant OKF bundle (rac export --okf), so you get the interchange for free and keep the enforcement OKF leaves out (ADR-048, ADR-049).
| Dimension | Lore / RAC | OKF (v0.1 Draft) |
|---|---|---|
| Validation time | Write-time: rac validate and rac relationships --validate fail CI before the knowledge lands |
Read-time: consumers "MUST NOT reject a bundle" for missing optional fields or unknown type, and "MUST tolerate broken links" |
| Links | Typed ## Related structural references, resolved and validated — broken, ambiguous, superseded-target, and unsupported edges are errors |
Untyped: the relationship kind "is conveyed by the surrounding prose, not by the link itself" |
type field |
Five enumerated types that drive classification and validation | A free string — "Type values are not registered centrally" |
| Cross-artifact checks | Referential integrity, status-consistency, edge-legality — enforced deterministically in CI | None defined; consumption is permissive by design |
| Maturity | Governed corpus; CLI and MCP output is a stability-tested contract | "Version 0.1 — Draft"; a single-vendor (Google Cloud) initiative |
| Interoperability | rac export --okf emits a conformant OKF bundle |
The shared carrier RAC writes |
How Lore earns trust
Lore asks you to trust it with your product knowledge, so it holds itself to the same standard it applies to your repository:
- The MCP server is read-only by construction. It cannot create, modify, or delete files in your repo — enforced in code and verified by tests, not by convention.
- No AI in the core. Retrieval is deterministic: the same repo state and the same query always return the same result. The reasoning is your agent's job; Lore's job is to hand it the facts.
- It dogfoods itself. Lore's own planning corpus under
rac/is validated by RAC in CI — if the tool's rules break the tool's own artifacts, the build fails. - Output is a contract. Golden tests pin CLI and MCP output; any change to what the tools return is reviewed as a product change.
- Telemetry is opt-in twice over. Local recording needs an explicit
--telemetryflag and never includes your arguments or repository content. Remote sharing is a separate, explicit consent (rac telemetry on, or one honest question atrac init): one anonymous daily ping — a random install id, the version, and an active-repo count — never paths, queries, or content.rac telemetry statusshows exactly what is shared, the network surface is a single readable module, and ADR-041 records the decision.
Sharing the corpus (the Portal)
Agents read your lore over MCP; people get the Portal — a single self-contained HTML file of the whole corpus that opens from file:// with zero network requests. Attach it to a release, send it to a stakeholder, open it on a plane.
rac export rac/ # canonical JSON to stdout
rac export rac/ --html --out lore-export.html # the Portal, one file
The JSON payload is a stable contract (artifacts with ids, aliases, status, rendered bodies; relationships as edges) for anyone building their own viewer. The Portal ships search, type/status filters, and citation cross-links out of the box.
Lore is early and evolving quickly. The MCP server ships today; feedback from teams running agents in anger is exactly what shapes what comes next. Contributions, ideas, and experiments welcome — see CONTRIBUTING.md.