Release Candidate · v0.9-rc1

FAI Protocol

The declarative context-wiring standard for AI primitive orchestration.

Model-agnostic · Framework-agnostic · Cloud-agnostic · MIT

In one paragraph

The FAI Protocol defines a JSON-based manifest format (fai-manifest.json) for declaring how AI primitives — agents, instructions, skills, hooks, workflows, plugins, tools, prompts, and guardrails — are wired together into a coherent, evaluated, deployable system. MCP standardised tool-calling. AG-UI standardised rendering. FAI standardises context-wiring + determinism + governance.

Read the spec (latest)

Full specification on GitHub →

12 sections + appendices. Manifest format, context resolution, path resolution, companion fai-context.json, conformance levels, security, versioning.

Versioned permalink

v0.9-rc1 immutable tag →

Cite this URL in papers, contracts, conformance claims. Content at this URL never changes.

Companion artefacts

JSON Schemas →
Formal Draft-07 schemas for fai-manifest, fai-context, plugin, agent, instruction, skill, hook, and the marketplace registry (v1.1).
VERSIONING.md →
Semver policy across spec, schemas, manifest format, conformance suite, and marketplace registry. 90-day deprecation rule. Path to v1.0 GA.
Examples →
10 minimal copy-paste-ready fai-manifest snippets, one per feature.
Conformance suite →
L0 = the 5 foundational tests every conformant runtime must pass. L1–L5 land with v1.0 GA. Run with npm run test:conformance or npx fai conformance.
Reference implementation (FAI Engine) →
6 cooperating modules: manifest-reader, context-resolver, primitive-wirer, hook-runner, evaluator, mcp-bridge. Node.js, MIT.

Public RFC — your input shapes v1.0

v0.9-rc1 is feature-complete for v1.0. We are now in the public comment period. If you build agents, framework runtimes, eval-tooling, or anything that touches AI primitives — read the spec and join the conversation.

Is the wiring model right?
Are the 9 primitive types complete?
Are the L0–L5 conformance bars set right?
Does the manifest schema feel natural to author by hand?
What prior art should we credit, learn from, or align with?

Join the RFC discussion →Read the RFC workspace →

Endorsements — voices backing v1.0

Endorsement window open

The v1.0 GA criteria require three or more public endorsements from distinct communities — at least one each from framework, LLM-provider, and eval-tooling perspectives. As of the open RFC, we are actively soliciting reviewers who would be willing to be quoted publicly when v1.0 ships.

If you build with agentic AI in production and your concerns are addressed in the spec, we'd be honoured to attribute your endorsement publicly.

Express endorsement interest →

Primitive types

(L0–L5)

Conformance levels

WAF pillars enforced