Gavio — Python SDK

AI request runtime and inspector for production systems. PII protection, audit trails, runtime events, reliability, cost intelligence, policy packs, production trust packages, and provider adapters as composable interceptors.

gavio sits between your application and any LLM provider. The same request passes through a pre/post interceptor chain — PII redaction, retries, caching, budgets, audit logging, tool runtime, runtime events, runtime context — before and after the provider call. Same API in Python, Java, and JavaScript, enforced by shared cross-SDK test vectors.

Part of the Gavio project. MIT licensed.

Install

pip install gavio            # zero mandatory dependencies
pip install gavio[redis]     # + distributed cache backend
pip install gavio[dev]       # + pytest, ruff, mypy

Requires Python 3.10+.

Quick start (dev mode — no API key, no network)

import asyncio
from gavio import Gateway
from gavio.interceptors.pii import PiiGuard

gw = (
    Gateway.builder()
    .dev_mode(True)                 # MockProvider + stdout audit
    .use(PiiGuard())                # redact PII before it leaves the process
    .build()
)

async def main():
    resp = await gw.complete(
        messages=[{"role": "user", "content": "Email jan@example.com about NL91ABNA0417164300"}],
        agent_id="demo",
    )
    print(resp.content)             # PII restored in the reply
    print(f"cost=${resp.cost_usd:.6f} latency={resp.latency_ms}ms")
    print("pii types:", resp.audit.pii_entity_types)

asyncio.run(main())

Real providers

from gavio import Gateway, Provider
from gavio.interceptors.pii import PiiGuard
from gavio.interceptors.audit import AuditInterceptor
from gavio.interceptors.reliability import RetryInterceptor, TimeoutPolicy

gw = (
    Gateway.builder()
    .provider(Provider.ANTHROPIC)          # reads ANTHROPIC_API_KEY
    .model("claude-sonnet-4-6")
    .use(PiiGuard(sensitivity="strict"))
    .use(AuditInterceptor(sink="stdout://"))
    .use(TimeoutPolicy(timeout_seconds=30))
    .use(RetryInterceptor(max_attempts=3))
    .build()
)

resp = await gw.complete(messages=[{"role": "user", "content": "Hi"}])

OpenAI, Gemini, Azure OpenAI, OpenRouter, and Ollama adapters work the same way (Provider.OPENAI, Provider.GEMINI, Provider.AZURE_OPENAI, Provider.OPENROUTER, Provider.OLLAMA) — switching providers is a config change, never an application change.

Streaming buffers the provider stream so post-interceptors (guardrails, PII restore, audit) run on the complete response before any chunk reaches you:

async for chunk in gw.stream(messages=[{"role": "user", "content": "Hi"}]):
    print(chunk, end="")

Tool Runtime

from gavio.interceptors.tool_runtime import ToolRuntimeInterceptor

gw = Gateway.builder().dev_mode(True).use(ToolRuntimeInterceptor()).build()

Tool Runtime reads metadata["tools"]["calls"], validates declared input/output schemas, checks result freshness, detects configured conflicts, and records provenance in ctx.tools["runtime"]. Tool Runtime v2 also understands definitions, permissions, approvals, records, and MCP metadata.

Embeddings run through the same pipeline — inputs are PII-scanned before the provider's embedding API is called:

resp = await gw.embed(["index this: contact jan@example.com"])
print(len(resp.embeddings))         # one vector per input, PII never left

The Inspector

An embedded, zero-dependency visualizer for the pipeline: live traces, per-interceptor waterfalls, PII redaction diffs, multi-agent call graphs, replay, RED stats, and a read-only production dashboard.

gw = Gateway.builder().dev_mode(True).inspect(True).build()
# open http://127.0.0.1:7411 and send a request

In production, write audits to a JSONL store and serve the dashboard from it:

from gavio.interceptors.audit import AuditInterceptor, JsonlSink
gw = Gateway.builder().provider(Provider.ANTHROPIC) \
    .use(AuditInterceptor(sink=JsonlSink("audit.jsonl"), hash_chain=True)).build()

gavio inspect --store audit.jsonl   # metadata mode: no content, no replay

Runtime export

from gavio import Gateway, JsonlRuntimeExporter, OtelSpanExporter

gw = (
    Gateway.builder()
    .dev_mode(True)
    .exporter(JsonlRuntimeExporter("runtime-events.jsonl"))
    .exporter(OtelSpanExporter("otel-spans.jsonl", service_name="checkout-api"))
    .build()
)

Runtime export (v1.1.0) writes metadata-safe JSONL events for integrations. The exporter strips messages, content, and diff by default, even when the local Inspector is in full capture mode. Observability + OTel (v1.3.0) maps the same stream into OpenTelemetry-style span JSON (F-OBS-07).

Self-hosted Control Plane

from gavio import Gateway

gw = (
    Gateway.builder()
    .dev_mode(True)
    .control_plane(
        "http://127.0.0.1:8787",
        runtime_key,
        "project:prod-support",
        fail_mode="open",
    )
    .build()
)

Control Plane support (v1.7.0) loads runtime config from an optional self-hosted server, caches the last successful config, and can fail open or closed during outages. The same surface is available as ControlPlaneClient and load_control_plane_config.

Production Trust Package

from gavio import build_production_trust_bundle, verify_production_trust_bundle

bundle = build_production_trust_bundle(
    bundle_id="trust-prod-support-2026-07-12",
    generated_at="2026-07-12T12:00:00Z",
    release={"version": "1.8.0", "tag": "v1.8.0"},
    runtime={
        "environment": "production",
        "policySource": "project:prod-support",
        "eventExportMode": "metadata_only",
    },
    audit_records=audit_records,
)

assert verify_production_trust_bundle(bundle).valid

Production Trust Package support (v1.8.0, F-TRUST-01) creates deterministic, metadata-only release evidence bundles for audit-chain, runtime-event, policy, benchmark, and document review.

Prompt Registry + Evals

from gavio import EvalSuite, PromptRegistry, PromptTemplate

registry = PromptRegistry([
    PromptTemplate(
        id="support.reply",
        version="2026-07-12",
        messages=[
            {"role": "system", "content": "You are concise."},
            {"role": "user", "content": "Reply to {{ customer }} about {{ topic }}."},
        ],
        required_variables=("customer", "topic"),
    )
])

report = await EvalSuite.from_dict({
    "id": "support-smoke",
    "cases": [{
        "id": "refund",
        "templateId": "support.reply",
        "variables": {"customer": "Avery", "topic": "refund"},
        "assertions": [{"type": "contains", "value": "refund"}],
    }],
}).run(registry, lambda _prompt, _case: "Avery refund approved")

Prompt Registry + Evals (v1.4.0) adds versioned prompt templates, metadata-only lineage, deterministic pass/fail reports, and SHA-256 output hashes instead of raw model output (F-EVAL-01/02).

What's inside

Every feature is an interceptor you compose explicitly — no hidden magic.

• Privacy & security — PII Guard with Email, IBAN (mod-97), BSN (11-proef), CreditCard (Luhn), Phone, IP, SSN scanners and redact/mask/tag/block + restore (F-SEC-01); secret/credential scanner (F-SEC-04); prompt injection guard (F-SEC-05); embedding call guard (F-SEC-10); Policy Pack manifests for core, FinTech, custom regex-rule packs, and the signed domain catalog with load/override/signature APIs (F-PACK-01/02/05).
• Reliability — retry with backoff (F-REL-01), provider fallback chain (F-REL-02), circuit breaker (F-REL-03), load balancing (F-REL-04), buffered streaming (F-REL-06), timeouts (F-REL-07).
• Caching — SHA-256 exact + semantic (cosine) cache with in-memory and Redis backends (F-CACHE-01/02/03/04).
• Cost & governance — per-request cost tracking (F-GOV-01), budget caps (F-GOV-02), rate limiting (F-GOV-03), per-role model policy (F-GOV-04), cost-optimiser routing (F-GOV-06), Cost Governance v2 budget policies, decisions, reports, and gavio cost report (v1.2.0).
• Observability — audit-by-default with SHA-256 content hashes, never raw text (F-OBS-01), tamper-evident hash chain (F-OBS-02), multi-agent DAG tracing via agent_id/parent_trace_id (F-OBS-03), prompt lineage (F-OBS-04), Prometheus metrics (F-OBS-08), stdout + JSONL sinks.
• Prompt Registry + Evals — versioned templates, lineage-preserving render, deterministic eval cases, privacy-safe output hashes (F-EVAL-01/02).
• Runtime export — metadata-safe JSONL runtime events (F-EXP-01) and OpenTelemetry-style span JSON (F-OBS-07) for gateway, observability, and eval integrations.
• Control Plane — optional self-hosted runtime config with policy rollout, budget config, audit search, config snapshots, SDK cache fallback, and ControlPlaneClient (v1.7.0).
• Production Trust Package — metadata-only release evidence bundles with deterministic hashes, privacy checks, audit-chain evidence, runtime-event evidence, and document/control pointers (F-TRUST-01).
• Quality — guardrails with JSON-schema and regex validators (F-QUA-01/02), composite risk scoring (F-QUA-06).
• Inspector — dev-time visualizer (F-DX-09/10), agent call graphs and session views (F-OBS-10), trace replay (F-DX-11), PII-sanitized test-case export (F-DX-12), read-only production dashboard + gavio inspect CLI (F-DX-08).
• Developer experience — dev mode (F-DX-01), dry-run (F-DX-02), GavioTestKit (F-DX-03), OpenAI drop-in shim (F-DX-04), config-file gateway construction (F-DX-05).
• Providers — OpenAI, Anthropic, Gemini, Azure OpenAI, OpenRouter, Ollama, Mock.

See the documentation site, the Python guide, the runnable examples, and the CHANGELOG for version-by-version detail.

Tests

pip install -e ".[dev]"
pytest tests/unit -v
ruff check gavio