Open Standard · Apache 2.0 · Originated from AMADEQ Skills Protocol

OPEN

SKILLS

PROTOCOL

The vendor-neutral standard for packaging, routing, and safely executing skills in multi-persona AI systems. Real BM25. Real embeddings. Real cryptographic signing.

pip install osp-sdk ▶ Live Demo GitHub →

# minimal pip install osp-sdk # + BM25 routing + sentence-transformers pip install osp-sdk[ml] # full: +crypto +monitoring +fastapi pip install osp-sdk[full]

# OSP conformance CLI tool npm install -g @osp/check # run against your implementation npx osp-check --level core

# Go reference client go get github.com/open-skills-protocol/osp-go # import in your code // import "github.com/open-skills-protocol/osp-go/routing"

<1ms

Stage 0 routing latency

17/18

Conformance tests passing

D0–3

Graceful degradation states

ES256

JCS RFC 8785 signed responses

1,500+

AI avatars in production (ASP)

94.4% passing · latest run 2h ago

Architecture

STANDARD vs.
IMPLEMENTATION

OSP is the open standard. ASP (AMADEQ Skills Protocol) is the reference implementation — proving the standard at production scale with 1,500+ avatars.

Open Standard · Vendor-Neutral

OSP

Open Skills Protocol · github.com/open-skills-protocol

JSON Schemas, routing contracts, conformance test vectors, and governance rules. Owned by the community. No single vendor controls merge rights. Apache 2.0.

Neutral GitHub org — 2-of-N RFC approval required
SkillManifest, RoutingDecision, TraceEvent, SafetyDecision schemas
4-stage routing pipeline contract (Filter → Score → Rerank → Resolve)
D0–D3 degradation automaton with named transitions
Any language, any model, any vector store — zero requirements
Apache 2.0 (spec) + CC BY 4.0 (docs)

Reference Implementation · AMADEQ Project

ASP

AMADEQ Skills Protocol · amadeq.org

Production OSP implementation powering MetaHuman AI tutors, counselors, and companions. Python + Qdrant + sentence-transformers on RTX 5090 infrastructure.

First OSP Resilience conformant implementation
Real BM25 (Okapi) + batch sentence-transformers Stage 2
KL-divergence anomaly brake in safety pipeline
JCS RFC 8785 + ES256/EdDSA/HMAC signing & verification
Configurable activation keyword strategies per skill
psutil auto-degradation monitoring, hysteresis controller

ASP implements OSP exactly as Chrome implements HTML or Nginx implements HTTP — the standard outlives any single implementation. Every team adopting OSP promotes AMADEQ as its origin. Every conformant implementation proves the spec is implementable from its text alone.

Quick Start

FROM ZERO
TO ROUTING

The @skill decorator transforms any Python function into an OSP-compliant skill with a full manifest, BM25 keywords, and cryptographic signing.

hello.py — examples/hello.py (real file)

from asp import skill, serve # @skill transforms any function into an OSP skill @skill("greet", description="Say hello to someone", keywords=["hello", "greet", "hi"], risk_level="LOW") def greet(name: str = "World") -> str: """Greet someone by name.""" return f"Hello, {name}! 👋" @skill("calculator", description="Perform math calculations", keywords=["math", "calc", "calculate"], risk_level="LOW") def calculator(expression: str) -> str: return str(eval(expression)) serve() # → http://localhost:8080 + /_dashboard

// TypeScript SDK (in development) import { skill, serve } from "@osp/sdk" @skill({ id: "greet", description: "Say hello to someone", keywords: ["hello", "greet", "hi"], riskLevel: "LOW", version: "1.0.0" }) async function greet(name: string): Promise<string> { return `Hello, ${name}!` } serve({ port: 8080 }) // OSP JSON-RPC 2.0

# Direct routing API (asp.route JSON-RPC) import requests result = requests.post("http://localhost:8080/asp-rpc", json={ "jsonrpc": "2.0", "method": "asp.route", "params": { "query": "Say hi to Alex", "routing_conditions": {"skip_semantic": False} }, "id": 1 }).json() # Response: # { skill_ref: "greet", safety_clearance: "allow", # decision_stability: "semantic_supported", # trace_events: [...] }

Install SDK

pip install osp-sdk[ml] — includes BM25 routing + sentence-transformers for Stage 2 semantic reranking.

~30 seconds

Decorate your functions

@skill() auto-extracts parameter types, generates L0 manifest, registers in global registry. One decorator, full OSP compliance.

~5 minutes

serve() — zero config

Starts JSON-RPC 2.0 server on :8080. Built-in /_dashboard, /health, /skills. Full 4-stage pipeline activated automatically.

~1 line of code

osp check — validate conformance

Run the OSP Core conformance suite against your implementation. 20 routing vectors + 5 negative tests. Pass → earn your badge.

~15 minutes

Architecture

THREE-PLANE
SEPARATION

Click each plane to explore what data is allowed and forbidden between layers. OSP enforces strict data isolation — raw queries never reach the LLM.

Plane 1

Policy / OSP

Routing, Safety, Governance

Deterministic routing pipeline (Stage 0–3). Safety classifier. Degrade automaton.
Owns: skill_id, risk_level, routing_decision, safety_clearance, trace_id

skill_id risk_level safety_clearance trace_id raw_query llm_output

↕ skill_id + safety_clearance + input_ref

Plane 2

Generation / Reasoning

LLM, RAG, Embeddings

LLM inference, RAG retrieval, embedding generation. Receives sanitized skill context only.
Owns: generated_text, embeddings, reasoning_trace

sanitized_context skill_parameters generated_output routing_scores safety_internals other_skill_data

↕ tool_call + action_ref

Plane 3

Action / MCP

Tools, Resources, APIs

Tool execution via MCP protocol. File access, API calls, database queries.
Owns: tool_results, resource_data, execution_logs

tool_call action_ref tool_results routing_decision safety_clearance llm_internals

Core Architecture

4-STAGE
ROUTING PIPELINE

No generative LLM in the routing path. Every stage deterministic. Every failure emits a TraceEvent — silent degradation is a conformance violation.

STAGE 0

Filter

Predicate

Candidate Reduction

L0 catalog-lite lookup — 4KB max per record. Capability, mode, and context predicates. Bounded and non-blocking. All 1000+ skills scanned, subset passed forward.

on failure

Empty result → ROUTING_POOL_EMPTY TraceEvent emitted. Never silent.

STAGE 1

Score

BM25

Lexical Scoring

Okapi BM25 with pre-built IDF table across candidate corpus. Compiled regex at module level. LRU cache (256 entries). Escape hatch for early Stage 2 promotion.

on failure

Timeout → ROUTING_FALLBACK_LEXICAL. No retry on same backend+config (I2 invariant).

STAGE 2

Rerank

Cosine

Semantic Reranking

Batch sentence-transformers encode (all-MiniLM-L6-v2). One .encode() call for query + all candidates. Combined score: α·BM25 + β·Semantic (0.4/0.6). ANN supported.

on failure

Vector store down → STAGE2_EMBEDDING_TIMEOUT → ROUTING_FALLBACK_LEXICAL. Emit ANN_APPROXIMATE_RECALL if ANN used.

STAGE 3

Resolve

Graph

Conflict Resolution

IEEE 754 fp64 epsilon comparison (ε=1e-6). Risk-level priority (LOW < MEDIUM < HIGH). UTF-8 lexicographic tiebreak on skill_id. O(1) amortized mutual-exclusivity query.

on failure

CONFLICT_GRAPH_UNAVAILABLE → canonical fallback: lowest risk → highest score → skill_id asc.

NORMATIVE: Routing stages 0–3 MUST be deterministic. No autoregressive LLM calls permitted in routing path. Combined score formula: score = 0.4 · normalize(BM25) + 0.6 · cosine_similarity. Tie-breaking: ascending skill_id (lexicographical UTF-8 byte order). Safety check runs before Stage 1.

Interactive

ROUTING
PLAYGROUND

Watch the real OSP 4-stage pipeline execute with actual BM25 scoring logic and safety checks from the reference implementation.

osp-routing-simulator · asp_server/logic/routing.py

● LIVE SIMULATION

Sample Queries

Say hi to Alex

Math homework

Emotional support

YouTube analysis

Google Drive

SQL injection ⚠

Query

L0 Skill Registry (6 skills · metadata.yaml)

greet@2.1.0 LOW math-tutor@3.0.1 LOW psych-counsel@1.4.0 MED youtube.analyzer@1.0.0 LOW google.drive@1.0.0 LOW crisis-support@2.0.0 HIGH

← Select a query and run the pipeline
Watch BM25 scoring, safety check,
and conflict resolution in real time

Safety Pipeline

THREE-LAYER
SAFETY

Fail-closed by default. Runs before Stage 1. TF-IDF semantic classifier with 6 risk categories. KL-divergence anomaly detection for distribution shift.

Regex Pre-filters

SQL Injection + Command Injection

Compiled regex patterns for SQL injection (UNION SELECT, DROP TABLE, OR 1=1) and command injection (rm -rf, cat /etc/passwd, $(...)). Fastest layer — microsecond response. Returns PREFILTER_SQL_INJECTION or PREFILTER_COMMAND_INJECTION reason codes.

Semantic Classifier

TF-IDF + Cosine Similarity · scikit-learn

6 risk categories: JAILBREAK, PRIVACY, ILLEGAL, VIOLENCE, MANIPULATION, INTERNAL_STATE. TF-IDF vectorizer with ngram_range(1,2). Cosine similarity threshold: 0.15 = suspicious, 0.25 = block. Falls back to keyword matching if sklearn unavailable.

LAYER 3 · ANOMALY DETECTION

KL-Divergence Distribution Brake

Tracks lexical vs semantic score distributions across last 100 requests (bounded deque).
Computes D_KL(P||Q) between distributions. Threshold: 0.5.
If KL > 0.5 AND risk_level is HIGH/CRITICAL → CONSERVATIVE_BLOCK_APPLIED + fail-closed.
If classifier throws any exception → FAIL_CLOSED_TRIGGERED → SAFETY_CLASSIFIER_UNAVAILABLE returned.

Resilience

DEGRADATION
AUTOMATON

Drag the sliders to simulate infrastructure failures. Watch the D0–D3 state machine respond in real time with deterministic transitions.

Vector DB Health 100%

Safety Classifier 100%

Embedding Backend 100%

CPU Load 20%

Full Operation

Semantic + lexical · Full tools · Sync

▼ DEGRADE_TRANSITION

Reduced

Semantic (reduced K) · Tools allow-list · Sync

▼ DEGRADE_TRANSITION

Minimal

Lexical only · No tools · Sync

▼ DEGRADE_TRANSITION

Async / Critical

Async routing · Full semantics · TTL

Transition events will appear here…

Full Operation

CPU <50% · RAM <60%

Reduced Intelligence

CPU 50–80% · RAM 60–85%

Minimal / Strict

CPU 80–95% · RAM 85–95%

Critical / Load Shed

CPU >95% · RAM >95%

Positioning

WHY OSP,
NOT X

MCP handles tool discovery. A2A handles agent-to-agent communication. OSP fills the gap: full skill lifecycle inside multi-persona systems.

Capability	MCP	OpenAI Tools	LangChain	A2A Google	OSP ←
Multi-persona skill routing	—	—	Partial	—	✓ Core
Safety tiering (LOW/MED/HIGH)	—	—	—	—	✓ 3-layer
Graceful degradation (D0→D3)	—	—	—	—	✓ psutil FSM
Deterministic routing contract	—	—	—	—	✓ IEEE 754
Cryptographic response signing	—	—	—	—	✓ JCS+ES256
Conformance test suite (YAML)	—	—	—	—	✓ 18+ vectors
Skill lifecycle governance	—	—	—	—	✓ LifecycleGates
Keyword activation tuning	—	Partial	Partial	—	✓ Configurable
Tool/resource discovery	✓	✓	Partial	—	→ use MCP
Agent-to-agent communication	—	—	Partial	✓	→ use A2A

Skills Registry

BUILT-IN
SKILL CATALOG

Production skills from the ASP reference implementation. Each skill is a metadata.yaml conforming to the OSP L0 Catalog-Lite Record specification.

foundation

Content Summarizer

org.osp.foundation.summarize@1.0.0

Summarizes text, articles, or audio files into concise bullet points. Configurable output language via parameters.

LOW hybrid summarize, summary, report

integrations · google

YouTube Analyzer

org.antigravity.youtube.analyzer@1.0.0

Extracts transcripts and provides summaries of YouTube videos by URL or video_id parameter.

LOW hybrid analyze, youtube, transcript

integrations · google

Google Drive

org.antigravity.google.drive@1.0.0

List and search files in Google Drive. Supports complex query strings matching Google Drive API syntax.

LOW hybrid list, file, drive, search, find

integrations · google

Invoice Search

org.antigravity.google.search_invoices@1.0.0

Search and retrieve invoices from Google Drive with structured metadata extraction.

LOW hybrid invoice, receipt, payment

integrations · telegram

Telegram Agent

org.antigravity.telegram@1.0.0

Send messages and interact with Telegram chats and channels programmatically.

MED hybrid telegram, send, message

contribute

Add Your Skill

org.yourname.skill@x.y.z

Create metadata.yaml following L0 Record spec. Submit PR to open-skills-protocol/registry.

→ osp new skill my-skill

Progressive Disclosure

SKILL LOADING
FUNNEL

Skills load progressively: L0 for catalog-level filtering (4KB budget), L1 for scoring candidates, L2 for full execution. Click each level to explore.

Catalog-Lite

≤4KB per record · Preloaded at boot

847+

▼ Stage 0–1 filter + BM25 score

Scoring Metadata

Keywords · Schema hash · Conflict edges

10–100

▼ Stage 2 rerank + resolve

Full Manifest

Tool schemas · Execution config · Auth

1–3

L0 — Catalog-Lite Record
Minimal metadata for fast filtering. Fields: skill_id, display_name, version, keywords[], description (≤200 chars), risk_level, activation_phrases[].
Budget: ≤4KB per record (JSON). Preloaded into memory at node startup for O(1) predicate filtering.
Schema: L0Record.schema.json (JSON Schema draft 2020-12)

Developer Tools

OSP CLI

Scaffold skills, run conformance checks, validate manifests, and monitor your deployment from the terminal.

osp new skill <name>

→

Scaffold metadata.yaml + scripts/tools.py + skill.md in seconds

osp new agent <name>

→

Create agent.yaml with skill dependencies and model config

osp check

→

Run OSP Core conformance suite (20 routing + 5 negative tests)

osp check --level safety

→

Safety tier: +10 safety cases, +5 adversarial samples

osp manifest validate

→

Validate metadata.yaml against OSP L0 Record JSON Schema

asp dev hello.py

→

Start ASP server in dev mode with hot reload + dashboard at /_dashboard

terminal · osp check output

$ osp check --level core

══════════════════════════════════════════

OSP Conformance Harness — 18 tests

══════════════════════════════════════════

✓ [RT-001] Basic Lexical Routing

✓ [RT-002] Empty Query Rejection

✓ [RT-003] Empty Pool Escalation

✓ [RT-004] Escape Hatch Override

✓ [RT-005] BM25 Best Match — got: org.test.weather

✓ [RT-006] Skip Semantic Condition

✓ [RT-007] Trace Events Present

✓ [RT-008] IEEE 754 Epsilon

✓ [SAF-001] SQL Injection Blocked

✓ [SAF-002] Command Injection Blocked

✓ [SAF-003] Safe Query Passes

✓ [SAF-004] Violence Detection

✓ [SAF-005] KL-Divergence Zero for Identical

✓ [DEG-001–004] D0/D1/D2/D3 States

✓ [MDL-001–004] Pydantic Models

✗ [MDL-001] Pydantic Models — pydantic not installed

──────────────────────────────────────────

✅ Passed: 17 ❌ Failed: 1 Total: 18

Score: 17/18 (94.4%) · OSP Core level

📄 Report: conformance_report.json

Error Taxonomy

CANONICAL
TRACE EVENTS

Every routing decision, safety block, and degradation transition emits a structured TraceEvent. All codes from the reference implementation.

Code	Stage	Description
ROUTING_DECISION_FINAL	3	Routing resolved successfully. skill_ref is set.
STAGE1_LEXICAL_MATCH	1	BM25 scoring completed. context includes latency_ms.
STAGE1_IDENTICAL_SCORES	1	Multiple candidates share identical BM25 score. Stage 2 required.
STAGE1_NO_MATCHES	1	All BM25 scores are 0. Fallback to first candidate with approximate=true.
STAGE2_EMBEDDING_GENERATED	2	Batch encoding successful. Cosine similarity computed.
STAGE2_EMBEDDING_TIMEOUT	2	Sentence-transformer timed out. Falls back to BM25-only ranking.
STAGE2_SEMANTIC_THRESHOLD_MET	2	Best semantic score ≥ 0.7. High confidence routing.
ANN_APPROXIMATE_RECALL	2	ANN backend used. recall_unknown: true if offline eval unavailable.
STAGE3_CONFLICT_DETECTED	3	Tied combined scores detected. Tiebreak procedure initiated.
STAGE3_TIE_BREAK_SKILL_ID	3	UTF-8 lexicographic tiebreak applied. tie_break_applied: true.
CONFLICT_GRAPH_UNAVAILABLE	3	Conflict graph offline. Canonical fallback chain used.
SAFETY_CHECK_PASS	safety	All 3 safety layers passed. Routing proceeds.
PREFILTER_SQL_INJECTION	safety	SQL injection pattern detected. refusal: true.
SEMANTIC_JAILBREAK_ATTEMPT	safety	TF-IDF classifier matched JAILBREAK category. Blocked.
ANOMALY_DETECTED	safety	KL-divergence > 0.5. Distribution shift detected.
FAIL_CLOSED_TRIGGERED	safety	Classifier threw exception. Fail-closed: request blocked.
DEGRADE_TRANSITION	degradation	D-state changed. from/to included in context.
SAFE_FALLBACK_SERVED	fallback	D3 active or routing failed. SafeFallbackResponse returned.
ROUTING_POOL_EMPTY	1	No candidate skills provided. safety_clearance: escalate.
CACHE_HIT	0	LRU cache hit. Identical query+candidates routed from cache.

Conformance

FOUR LEVELS,
ONE SUITE

Conformance is defined by passing official YAML test vectors — not self-attestation. Current status: ASP reference implementation 17/18 (94.4%).

OSP Core

Core

Routing vectors20

Negative tests5

Safety cases—

Degrade drills—

Adversarial—

OSP Core Conformant

OSP Safety

Safety

Routing vectors20

Negative tests8

Safety cases10

Degrade drills—

Adversarial5

OSP Safety Conformant

OSP Resilience

Resilience

Routing vectors20

Negative tests10

Safety cases10

Degrade drills15

Adversarial5

OSP Resilience Conformant

OSP Enterprise

Enterprise

Routing vectors50

Negative tests15

Safety cases20

Degrade drills15

Adversarial10

OSP Enterprise Conformant

ASP reference impl · last run 2h ago ·

17 passing

1 failing — pydantic not in env

Level claimed: OSP Resilience

Supply Chain

TRUST REGISTRY
EXPLORER

Explore the append-only Transparency Log. Register skills, simulate revocations, and detect split-view attacks on the Merkle tree.

🏛

OSP-TL Root

root_hash: a7f3...c912

TRUSTED

📦

org.amadeq.financial-advice@2.1.0

sha256: d4e5...f678

ACTIVE

📦

org.amadeq.greet@2.1.0

sha256: b2c3...a901

ACTIVE

📦

org.amadeq.knowledge-general@1.0.0

sha256: e5f6...b234

ACTIVE

Registry operations will appear here…

Governance

LIFECYCLE
GATES

Formal skill promotion with STATIC gates (100% blockers) and BEHAVIORAL gates (statistical drift detection). No other protocol does this.

DEV

local

→

STAGING

gates

→

CANARY

5% traffic

→

PROD

100%

STATIC Jailbreak test suite

STATIC SQL injection vectors

STATIC Signature verification

BEHAVIORAL Personality drift

BEHAVIORAL Hallucination rate

BEHAVIORAL Edge-case adversarial

Strategic Vision

ADOPTION
ROADMAP

A realistic, phased plan — not vaporware. Standards are born from independent implementations that agree on contracts.

YOU ARE HERE

Phase 1

De-facto Standard

6–12 months

✓ Position Paper v0.4.0

✓ 9 JSON Schemas published

✓ 86+ conformance vectors

✓ Reference implementation (ASP)

○ 2–3 independent implementations

○ Community review cycle

Phase 2

Community Standard

12–24 months

○ OSP Working Group formed

○ Schema stabilization (v1.0)

○ Multi-vendor conformance

○ Backward compatibility guarantee

○ DID federation (Phase 2 trust)

Phase 3

Formal Standard

24+ months

○ W3C / OASIS submission

○ Multiple independent adoptions

○ Formal versioning + ISO track

○ Post-quantum algorithm migration

"Standards are not born from specifications. They are born from independent implementations that agree on contracts."

Adopt OSP

FROM SPEC
TO BADGE

Estimated 4–8 hours for an experienced team to reach OSP Core conformance. No vendor lock-in. No API keys. Pure open standard.

Read osp-core-spec.md

Vendor-neutral specification. Zero implementation-specific references. Understand the 4-stage routing contract and D0–D3 automaton.

~15 min read

Implement Schemas

SkillManifest + RoutingDecision + TraceEvent + SafetyDecision. Any language. Any framework. Pydantic models available as reference.

~2–4h for core

Run Conformance

osp check — runs 20 routing vectors + 5 negative tests. Produces JSON report. Tests are YAML-defined, language-agnostic.

~1h for Core level

Submit & Display Badge

Open PR with conformance_report.json. Pass → join Adopters list + display your level badge in your README.

Community review 48h

OSP Adopters

1 independent · v0.3.x · growing

AMADEQ (ASP)

Python · sentence-transformers · Qdrant · RTX 5090

OSP Resilience · Reference Impl · 17/18

Your implementation here
osp check → open a PR →

Second independent impl
Needed for Working Group formation

OPEN

SKILLS

PROTOCOL

STANDARD vs.
IMPLEMENTATION

FROM ZERO
TO ROUTING

THREE-PLANE
SEPARATION

4-STAGE
ROUTING PIPELINE

ROUTING
PLAYGROUND

ROUTING
SIMULATOR

THREE-LAYER
SAFETY

SAFETY
SIMULATOR

DEGRADATION
AUTOMATON

WHY OSP,
NOT X

BUILT-IN
SKILL CATALOG

SKILL LOADING
FUNNEL

OSP CLI

CANONICAL
TRACE EVENTS

FOUR LEVELS,
ONE SUITE

CONFORMANCE
RUNNER

TRUST REGISTRY
EXPLORER

ERROR
TAXONOMY

SEMANTIC
ANOMALY BRAKE

DELIVERY
CONTRACT

LIFECYCLE
GATES

LIVE
TRACE STREAM

ADOPTION
ROADMAP

ASP IN
PRODUCTION

FROM SPEC
TO BADGE

OPEN SKILLS PROTOCOL

STANDARD vs.IMPLEMENTATION

FROM ZEROTO ROUTING

THREE-PLANESEPARATION

4-STAGEROUTING PIPELINE

ROUTINGPLAYGROUND

ROUTINGSIMULATOR

THREE-LAYERSAFETY

SAFETYSIMULATOR

DEGRADATIONAUTOMATON

WHY OSP,NOT X

BUILT-INSKILL CATALOG

SKILL LOADINGFUNNEL

OSP CLI

CANONICALTRACE EVENTS

FOUR LEVELS,ONE SUITE

CONFORMANCERUNNER

TRUST REGISTRYEXPLORER

ERRORTAXONOMY

SEMANTICANOMALY BRAKE

DELIVERYCONTRACT

LIFECYCLEGATES

LIVETRACE STREAM

ADOPTIONROADMAP

ASP INPRODUCTION

FROM SPECTO BADGE

OPEN

SKILLS

PROTOCOL

STANDARD vs.
IMPLEMENTATION

FROM ZERO
TO ROUTING

THREE-PLANE
SEPARATION

4-STAGE
ROUTING PIPELINE

ROUTING
PLAYGROUND

ROUTING
SIMULATOR

THREE-LAYER
SAFETY

SAFETY
SIMULATOR

DEGRADATION
AUTOMATON

WHY OSP,
NOT X

BUILT-IN
SKILL CATALOG

SKILL LOADING
FUNNEL

CANONICAL
TRACE EVENTS

FOUR LEVELS,
ONE SUITE

CONFORMANCE
RUNNER

TRUST REGISTRY
EXPLORER

ERROR
TAXONOMY

SEMANTIC
ANOMALY BRAKE

DELIVERY
CONTRACT

LIFECYCLE
GATES

LIVE
TRACE STREAM

ADOPTION
ROADMAP

ASP IN
PRODUCTION

FROM SPEC
TO BADGE