For AI & developers

Use practical knowledge without throwing away its evidence boundary.

Brali is a versioned practical-knowledge layer for humans and agents. The same canonical records power readable protocols, Agent Skills, a static API, local MCP tools, datasets, reference demos and evaluation cases. Evidence state, provenance and stable identity stay attached.

Start in 60 seconds

Try the useful part before learning the architecture.

You do not need to clone the repository or read the ontology first. Start with one of these four paths.

No setup

Ask Brali

Type a practical question, inspect the matched Topic, trusted Protocols, evidence state and provenance, then copy the bounded packet into an agent.

Open Query Playground

Agent Skill

Preview one Brali router

The router selects only from the trusted recommendation layer. Preview it before installation:

gh skill preview Brali-LifeOS/brali-lifeos.github.io brali-life-os

Then install it for the host and scope you actually use. Example for Claude Code:

gh skill install Brali-LifeOS/brali-lifeos.github.io brali-life-os --agent claude-code --scope user

Browse the complete one-skill-per-hack library →

Reproduce

Run the 50-case evaluation

Inspect the source questions, expected trust boundaries and generated results. The suite evaluates Brali retrieval and grounding; it does not pretend to benchmark an unpinned LLM.

Open Agent Evaluation →

Install

Complete AI Agent Skills library

Every Brali hack gets a portable SKILL.md. Trust mode determines whether it is usable, review-required or restricted-reference. New hacks receive skills automatically on the next build.

Open Agent Skills library →

Try it

Ask Brali

Run a zero-install evidence-aware query in your browser, then copy the grounded packet into your agent.

Open Query Playground →

Choose the smallest integration surface that fits

If an agent supports Agent Skills, start with the brali-life-os router or one bounded skill. For applications that need structured search, use API v1. Use the local MCP server when MCP fits your development environment. Do not copy the complete corpus into every prompt by default.

Recommended retrieval flow

  1. Identify the user's concrete Topic.
  2. Search /api/v1/search.json and resolve a canonical ID.
  3. Retrieve the Topic, Hack, Protocol or generated Agent Skill.
  4. Read evidence metadata and skill mode before making evidence-like or practical claims.
  5. For research-derived content, inspect the Evidence Decision: supported claim, unsupported claims, limitations and editorial outcome.
  6. Keep evidence state, skill mode, source scope, canonical ID and Brali URL downstream.
  7. Return no answer instead of inventing coverage.

Agent Skills

Every Brali hack is generated as a stable SKILL.md artifact under /skill-packs/. The complete machine library is /skill-packs/library.json; the smaller trusted/recommendation catalog is /skill-packs/catalog.json. Reviewed/practical records become usable skills. Pending-review records become review-required skills. Restricted records become non-operational restricted-reference skills.

Do not erase uncertainty

Trust metadata is part of the content, not decoration. Only reviewed and eligible practical content belongs in normal trusted recommendations. pending-review and restricted states must remain visible. A source record, a skill file or a search-indexable review page is not automatically a recommendation.

MCP

The repository includes a read-only MCP server in /mcp/ with search_knowledge, get_hack, get_protocol, get_evidence, list_topics, and get_related. It reads the same generated API files rather than maintaining a second knowledge model. It currently runs locally over stdio; Brali does not claim a hosted remote MCP endpoint.

Multilingual retrieval

English remains the canonical editorial language. The identity layer can add localized labels and synonyms while keeping one canonical entity ID.

Evaluation

Agent Evaluation Suite currently uses 50 practical questions to compare a no-knowledge grounding control, lexical retrieval and structured Brali retrieval, including safety, provenance and evidence-boundary cases. The source cases and generated report are public so the result can be inspected rather than trusted as a marketing number.

Quality and maintenance

Use the retrieval benchmark, the migration queue, and evidence metrics. Research discovery checks Crossref and Europe PMC, while discovery metadata remains unreviewed until actual source review.

Discovery & verification

Brali publishes separate surfaces for crawler policy, project-level factual claims, machine-surface inventory and reusable attribution facts. Crawler permission is not evidence of indexing, a project fact is not scientific evidence, and an internal evaluation is not proof of downstream adoption.

Machine-readable files

Reproducible use

The first stable dataset baseline is data-v1.0.0. Pin a release for research or long-lived integrations instead of following main. The bundle includes canonical datasets, the API v1 payload, checksums, citation/license metadata, policies, and version-specific release notes with known limitations.

Contribute or partner

Use Partnerships for corrections, sources, taxonomy work, research collaboration, integrations, localization, or licensing.