Context layer for AI development

Stop re-explaining yourself to your agents.

Decisions made in Claude don't reach Cursor. Libra fixes that — automatically writing your specs, decisions, and rules into your repo so every agent starts from the same page.

Get started View on GitHub

Decision (in Claude)

We're not building a REST API. Everything client-facing goes through tRPC — we want end-to-end type safety and we never want to hand-write a schema again. If something can't be done in tRPC, we figure out why before reaching for REST.

Opus 4.5

Libra writes it to your repo

docs/decisions/adr-004-trpc.md

ADR-004 — Use tRPC for all client-server communication

Status: Active

Decision

All client-facing API calls use tRPC. No REST endpoints.

Rationale

End-to-end type safety without hand-written schemas. Eliminates a category of runtime bugs at zero cost.

Consequences

New API surface must go through tRPC router in /server/router. REST only permitted if a third-party integration requires it, which requires a follow-up ADR.

Libra enforces your decisions

.libra/no-rest-endpoints.rule.sh

#!/bin/bash
# ADR-004 — Use tRPC for all client-server communication
# docs/decisions/adr-004-trpc.md

if grep -r "app\.\(get\|post\|put\|delete\)" src/ --include="*.ts"; then
  echo "❌ REST endpoint detected. All API calls must use tRPC."
  echo "   See docs/decisions/adr-004-trpc.md"
  exit 1
fi

echo "✅ No REST endpoints found."

Your agents read this before they write a single line of code.

How it works

Claude / ChatGPT

ideation & planning

Libra MCP

writes specs, decisions & rules to your repo

Claude / Cursor

reads context, writes code

Libra Skills

updates docs automatically

Pull Request

latest docs committed, review & merge

Libra App

./libra rules run in CI

Claude / ChatGPT

latest context always available

Getting started

Step 1 — Libra Skills

Start documenting automatically

Install the Libra skills plugin from the repo into Cursor or Claude Code. Decisions you make mid-conversation get written back to /docs in your repo.

github.com/libra-mcp/libra-skills

What this closes: Claude writes specs and ADRs into your repo as you plan. Cursor reads them before it writes any code. You never re-explain the same decision twice.

Step 2 — Libra MCP

Connect your chatbot to your docs

Add the Libra MCP server to Claude Desktop or any MCP-compatible chatbot. Your chatbot can read and write to your repo docs mid-conversation — specs, decisions, anything. Sign in to connect your repos and get set up.

Step 3 — Libra App

Enforce your rules in CI

Install the Libra GitHub App on your repo. Every PR is automatically checked against your .libra/rules.md before anything merges that contradicts what you decided.

Coming soon. Install steps will be available when the app is released.