Skip to main content

CRP SDK Reference

The CRP SDK is the developer-facing surface for the entire Context Relay Protocol. It is designed around progressive disclosure: the common path is five lines of code, and every advanced capability across all 48 CRP specs is reachable through the same client. Every capability is AIUC-1 aligned where applicable, from real-time input filtering to tamper-evident audit export.

• Level 0 - Governance: drop-in replacement for OpenAI/Anthropic clients; every call becomes governed, risk-scored, and audited.
• Level 1 - Quality: give CRP knowledge, then ask questions with grounded, cited answers.
• Level 2 - Control: depth, tools, safety profiles, reasoning transparency, and output shaping.
• Level 2+ - Agent SDK (SPEC-059): declarative agents with crp.Agent - declare tools + policy + model and let the protocol run the positioned loop.
• Level 3 - Infrastructure: raw protocol headers, audit export, compliance evidence, conformance tests, amplification, and multi-agent safety budgets.

This reference is the canonical map from installation to every public capability. For task-oriented tutorials, see the SDK Guide and Quickstart.

Implementation status

The CRP Python SDK is shipping with v6.0.0. CRPv6 Agent SDK (crp.Agent) is implemented, tested, and proven live with local SLMs. Advanced features such as amplification and agent-dispatch safety budgets are opt-in and declared upfront.

---

Install

pip install crprotocol            # core SDK
pip install crprotocol[full]      # + local model support, all ingest formats

import crp
print(crp.__version__)            # 6.0.0

Provider-agnostic

CRP does not replace your provider SDK. For hosted models install openai or anthropic; for local models point at LM Studio, Ollama, or vLLM.

---

Level 0 - Governance

0.1 Drop-in with the OpenAI SDK

Point any OpenAI-compatible client at the CRP Gateway. Nothing else changes.

from openai import OpenAI

client = OpenAI(
    api_key="crp_gw_...",
    base_url="https://your-gateway.example/v1",
)

response = client.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "Hello"}],
)

Gateway availability

Managed cloud at gateway.crprotocol.io is on the waitlist. You can also self-host the Gateway runtime locally or in your own infrastructure.

0.2 Native client

client = crp.SDKClient()                           # sane defaults, auto provider
client = crp.SDKClient(model="gpt-4o-mini")        # pick a model
client = crp.SDKClient(model="local:llama-3.1-8b") # local via LM Studio / Ollama
client = crp.SDKClient(api_key="crp_gw_...")       # hosted gateway

Legacy client

crp.Client is the legacy CRPOrchestrator. New code should use crp.SDKClient (also importable as crp.CRPClient).

0.3 Governance summary on every result

r = client.complete("Summarise the EU AI Act")

r.text                  # the output string
r.crp.risk              # LOW | MEDIUM | HIGH | CRITICAL
r.crp.grounded          # bool
r.crp.fabrications      # count of unsupported invented claims
r.crp.compliant         # bool - passed configured compliance checks
r.crp.audit_url         # tamper-evident audit deep link
r.crp.chain_valid       # bool - HMAC provenance chain intact
r.crp.injection_detected # bool - prompt injection markers found
r.crp.pii_detected      # bool - personal data detected in output
r.crp.safety_budget_remaining # 0.0–1.0

0.4 Enforcement, not just reporting

Policy is enforced automatically. If a call violates a halt-on rule, CRP raises a first-class exception that carries the evidence.

try:
    r = client.complete("...")
except crp.SecurityInvariantError as e:
    print(e.message)       # e.g. "CRITICAL hallucination risk"
    print(e.code)          # 1011
    print(e.details)       # evidence dict

---

Level 1 - Quality

1.1 ingest() - give CRP knowledge

client.ingest("./docs/")                       # directory (recursive)
client.ingest("manual.pdf")                    # single file
client.ingest("https://example.com/guide")     # URL
client.ingest("Raw text as a string")          # string
client.ingest(["a.md", "b.pdf", url, text])    # any mix

ingest() returns None. Extraction runs in the background; use client.knowledge.health() or client.storage.overview() to inspect progress.

1.2 ask() - the happy path

a = client.ask("Write a complete guide to our deployment process")

a.text                 # full grounded output
a.quality              # S | A | B | C | D
a.sources              # [{title, doc_id, used_facts, relevance_score}]
a.complete             # bool - covered the whole task
a.crp                  # same governance summary as Level 0

Behind ask() the SDK runs task classification, CDR/CDGR retrieval, multi-window continuation with CSO relay, DPE verification, and assembly. The developer sees none of it.

1.3 Streaming

Use client.stream() for token-by-token output with continuation-window events. Async streaming (aask_stream) is not yet implemented.

for event in client.stream("Explain the CAP theorem"):
    if event.event_type == "token":
        print(event.data, end="", flush=True)

See the Streaming guide for event types and examples.

1.4 Conversations

client = crp.SDKClient()
r1 = client.complete("What is the EU AI Act?")
r2 = client.ask("Which articles apply to high-risk systems?")

# Both calls share the same session and CKF state.
s = client.session()
print(s.id, s.fact_count, s.window_count)

---

Level 2 - Control

2.1 Depth

client.ask("...", depth="quick")       # short answer, low token cap
client.ask("...", depth="standard")    # balanced coverage
client.ask("...", depth="thorough")    # higher token cap, deeper envelope
client.ask("...", depth="exhaustive")  # continuation loop with stitching
client.ask("...", depth="auto")        # default

2.2 Tools

@client.tool
def get_metrics(service: str) -> dict:
    """Current CPU/memory for a service."""
    return fetch(service)

# Call the registered tool directly
metrics = client.call_tool("get_metrics", "payment-service")

# Or let the model invoke it during ask()
a = client.ask("Should we scale the payment service?")

Tool support requires a tool-capable provider

Automatic tool invocation uses the provider's generate_chat_with_tools() path. CustomProvider and other providers that do not report supports_tools() will skip the tool loop and fall back to standard dispatch.

2.3 Safety

client = crp.SDKClient(safety="balanced")   # default
client = crp.SDKClient(safety="strict")     # require approval, halt on injection/PII
client = crp.SDKClient(safety="permissive") # fully autonomous
client = crp.SDKClient(safety="research")   # informed logging, no halts

# Fine-grained
client = crp.SDKClient(safety={
    "profile": "strict",
    "alert_on_quality_below": "B",
})

# Runtime tuning
client.configure(safety.profile="strict")

Safety enforcement

Safety profiles are mapped to the orchestrator's human-oversight configuration. Gateway deployments enforce the same profile via the CRP-Safety-Policy header.

2.4 Reasoning transparency

a = client.ask("Design our database architecture")

a.decisions            # list of decisions / trade-offs if available
a.how_it_was_built     # simple human-readable construction summary
a.open_questions       # unresolved gaps

2.5 Application profile

Pass an existing application contract so CRP chooses the right provider, strategy, and provenance attestation.

from crp.core.app_profile import (
    ApplicationProfile,
    ProviderKind,
    ContextStrategy,
    build_profile_from_messages,
)

profile = ApplicationProfile(
    provider=ProviderKind.OPENAI,
    provider_model="gpt-4o-mini",
    context_window=128_000,
    context_strategy=ContextStrategy.RAG,
)
client = crp.CRPClient(app_profile=profile)

# Or derive it from existing messages/tools
profile = build_profile_from_messages(messages, tools=tools)
client = crp.CRPClient(app_profile=profile)

See the SDK Guide for the strategy mapping and manifest attestation details.

2.6 Knowledge control

client.knowledge.search("etcd")       # inspect what CRP knows
client.knowledge.query(...)           # CKF pattern query
client.knowledge.health()             # CKF health snapshot
client.knowledge.location             # "CKF"

2.7 Output shaping

Not yet implemented

format, max_words, style, and schema parameters are not currently implemented on ask().

---

Level 3 - Infrastructure

3.1 Raw protocol surface

r = client.complete("...")
r.raw_headers          # raw response headers, including CRP-* when present

Partial surface

Full DPE report (r.crp.dpe), raw CSO (r.crp.cso), and envelope details (r.crp.envelope) are not exposed on SDKClient responses today.

3.2 Audit & provenance

print(client.audit.summary())
print(client.audit.verify())         # (valid, broken_at_sequence)
export = client.audit.export()       # tamper-evident audit document

3.3 Compliance evidence

assessment = client.compliance.classify(
    intended_purpose="Customer support chatbot",
    processes_personal_data=True,
)
report = client.compliance.report()
controls = client.compliance.controls()

3.4 Conformance

Not yet implemented

crp.conformance.run() and crp.conformance.sqb() are not exposed on the SDK client.

3.5 Amplification (opt-in, async)

Not yet implemented

amplify, amplify_async, and client.amplify* methods do not exist.

3.6 Multi-agent

The client.agent namespace exposes multi-agent safety budgets, delegation checks, fan-in quality results, and oversight approvals:

budget = client.agent.budget()
print(budget.health)

High-level client.agent(role) task dispatch is not yet implemented.

---

Complete method map

SDKClient()
  .complete(prompt)              # single governed completion (Level 0)
  .stream(prompt)                 # synchronous streaming completion (Level 0)
  .ingest(source)                 # add knowledge (Level 1); returns None
  .ask(task)                      # happy path - full pipeline (Level 1)
  .tool / @client.tool            # register a tool (Level 2)
  .call_tool(name, ...)           # execute a registered tool (Level 2)
  .session()                      # live session view (Level 2)
  .knowledge.query()              # CKF pattern query (Level 2)
  .knowledge.search()             # semantic search (Level 2)
  .knowledge.health()             # CKF health snapshot (Level 2)
  .storage.overview()             # storage visibility (Level 2)
  .storage.fact_count()           # fact count (Level 2)
  .audit.export()                 # tamper-evident audit export (Level 3)
  .audit.verify()                 # HMAC chain verification (Level 3)
  .compliance.classify()          # EU AI Act risk classification (Level 3)
  .compliance.report()            # compliance status report (Level 3)

# Not yet implemented
  .ask_stream()                   # async streaming (planned)
  .agent(role)                    # multi-agent
  .amplify*                       # amplification
  .forget()                       # targeted erasure
  .last                           # most recent result object

---

Namespace map

Every advanced subsystem is reachable through a typed namespace on the client. Each namespace is documented in the API Reference.

Namespace	One-line description
client.activation	Activation modes and onboarding stages.
client.agent	Multi-agent safety budgets and oversight chains.
client.audit	Tamper-evident audit trail export and verification.
client.ckf	Contextual Knowledge Fabric: search, query, and inspect the fact graph.
client.compliance	Regulatory classification and compliance reports.
client.comply	Comply Gateway client, quota gate, no-code governance, and billing webhooks.
client.cso	Cognitive State Object relay across continuation windows.
client.events	Observability event bus for telemetry and logging.
client.extract	Document and tool-result extraction pipeline.
client.gateway	Gateway session, router, and key vault accessors.
client.headers	CRP HTTP header emission, parsing, and halt detection.
client.knowledge	Semantic search and CKF health.
client.observability	Audit logs, metrics, telemetry, and structured logging.
client.policy	Safety policy enforcement, profiles, and violation reporting.
client.providers	Provider discovery, adapters, and selection.
client.provenance	Claim grounding, fabrication detection, and DPE controls.
client.reasoning	Meta-learning and reasoning-scaffolding controls.
client.safety	Safety rules, checkpoints, and human-in-the-loop control plane.
client.scan	GitHub App, remediation proposals, and semantic code ingestion.
client.storage	Warm-state visibility and fact counts.
client.core	Orchestrator, session, DAG, config, and manifest ledger.
client.continuation	Continuation manager, document map, flow, voice, and residual gap.
client.envelope	Envelope builder, packer, reranker, CDR, and formatter.
client.state	Warm store, cold storage, snapshots, event log, and facts.
client.security	Audit trail, consent, RBAC, checkpoints, and encryption.
client.resources	Adaptive allocator, cost model, and resource manager.
client.advanced	Curator, feedback, meta-learning, and source grounding.
client.cli	Sidecar handler and startup helpers.
client.errors	Public exception classes.

---

Full coverage escape hatches

The namespace proxies above cover the most common subsystems. For the remaining 1,400+ classes and 700+ functions, the client exposes two dynamic accessors:

Accessor	What it reaches	Example
client.orchestrator	The live CRPOrchestrator instance and all its subsystems	client.orchestrator.dispatch(...)
client.modules	Any public class/function in any crp.* module	client.modules.envelope.cdr.cdr_rank(...)

These accessors are documented in the API Reference and keep the SDK in sync automatically as the codebase grows.

---

Design rules

1. Level 0/1 methods require zero protocol vocabulary. No "envelope", "window", "CDR", or "DPE" in signatures, return fields, or error messages.
2. Every return object degrades gracefully. r.crp.risk always exists, even in zero-CKF mode.
3. Errors state what + why + the one fix. No internal stack traces.
4. Defaults are the product. A developer who configures nothing gets safe, grounded output.
5. Autocomplete is the documentation. Level 2/3 capabilities are discoverable via IDE autocomplete.
6. The OpenAI-compatible path is truly drop-in. Swapping base_url changes nothing else.
7. Async mirrors sync. (Planned - async methods are not yet implemented.)