egisai 0.27.0

Native runtime governance SDK for AI agents — import egisai; egisai.init() and you're protected.

egisai — Runtime governance for AI agents

Production guardrails for Python AI applications. Install the SDK, call egisai.init(), and continue using OpenAI, Anthropic, Google Gemini, or plain HTTP clients as you do today—policy evaluation and audit logging wrap those calls automatically.

This document is the canonical SDK guide for PyPI and mirrors what we publish at docs.egisai.co.

---

Overview

Capability	What it means for you
Central policies	Operators configure rules in the EgisAI dashboard. The SDK loads them at runtime and refreshes them continuously—no redeploy to tighten controls.
Transparent integration	No proxy layer and no wrapper objects you must remember to use. Supported libraries are patched in-process when your application imports them after egisai.init().
Audit trail	Governed calls emit structured events to your org so teams can review verdicts, latency, and usage in one place.
Local-first sensitive checks	Pattern-based PII handling and other deterministic rules run entirely inside your process before traffic leaves your environment.

---

What you need

1. Python 3.11+
2. An EgisAI account and an SDK API key (dashboard → API Keys → create). Keys look like egis_live_….
3. The AI SDK(s) you already use (openai, anthropic, google-genai, …).

---

Installation

pip install "egisai[all]"

Optional extras (smaller installs):

pip install "egisai[openai]"
pip install "egisai[anthropic]"
pip install "egisai[google]"          # google-genai
pip install "egisai[google-legacy]"   # google-generativeai

Only frameworks present in your environment are activated at runtime. The google and google-legacy extras are independent and can both be installed when an application uses both SDKs.

---

Getting started

1. Initialize once per process

Call egisai.init() as early as possible in your application lifecycle (for example right after loading configuration). Use your SDK API key from the dashboard.

import egisai

egisai.init(
    api_key="egis_live_…",   # or set EGISAI_API_KEY in the environment
    app="customer-support-bot",
    env="production",
)

2. Use your LLM client normally

No changes to your calling convention—the SDK intercepts supported APIs after initialization.

OpenAI

from openai import OpenAI

client = OpenAI()
response = client.chat.completions.create(
    model="gpt-4.1",
    messages=[{"role": "user", "content": "Hello"}],
)

Anthropic

import anthropic

client = anthropic.Anthropic()
response = client.messages.create(
    model="claude-sonnet-4-20250514",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello"}],
)

Google Gemini

from google import genai

client = genai.Client()  # picks up GEMINI_API_KEY from the environment
response = client.models.generate_content(
    model="gemini-2.0-flash",
    contents="Hello",
)

3. Review activity

Open Dashboard → Requests to see governed calls, verdicts, and supporting metadata for your organization.

---

How governance fits your call path

Each governed call is evaluated in two phases—once before the model runs and once after it returns—so policies can intervene on either side independently.

1. Pre-model evaluation — Before the upstream model runs, the SDK applies your organization’s active policies (cached locally) to the prompt. Local deterministic rules — PII detection, regex denylists, model allowlists, prompt-size caps — run first; intent-oriented semantic_guard is consulted only if every local rule allowed the prompt.
2. Outcomes (pre-model) — A call may be allowed, sanitized (payload adjusted per policy, then forwarded), or blocked. Blocked calls never reach the provider when enforcement raises or returns a stub, depending on configuration (see below).
3. Post-model evaluation — When the model responds, output-side policies run against the assistant’s text, tool invocations, and connector targets, using the same two-phase split. Local rules run first; semantic_guard runs only if no local rule already blocked. A blocked response is suppressed before it reaches your code.
4. Telemetry — The audit event records each phase’s decision independently (prompt_decision, response_decision) so your dashboard can show exactly which side fired and which rules matched. Delivery is non-blocking.

Each policy carries a phase that selects which side it runs on:

Phase	When the rule fires
pre_model	Only on the prompt, before the model is called.
post_model	Only on the response, after the model returns.
both	On both sides where the rule type has signals to evaluate.

Operators choose the phase in the dashboard when they create or edit a rule, and every rule type accepts every phase. The engine evaluates each rule on whichever side it has meaningful signals for: text-content rules (PII detection, regex deny-lists, prompt-size caps, semantic guard) fire symmetrically on prompt and response; tool / shell / connector rules need response-side signals and silently no-op when an operator targets them on pre_model only. Older platform deployments that haven’t shipped the field yet behave as if every rule were both, preserving previous semantics.

Sensitive pattern detection intended to catch regulated data is performed locally so raw values are not sent to third-party models as part of governance — on both the prompt and the response. Intent-oriented policies (semantic_guard) are consulted only after the applicable local checks have run on the text that will be judged, and never run at all when a local rule has already refused the call. The same ordering applies to whichever side the operator scoped the rule to.

---

When a call is blocked

on_block	Behavior
"raise" (default)	Raises PermissionError if a policy blocks the call.
"stub"	Returns a framework-shaped refusal object so applications that cannot tolerate exceptions keep running; the refusal is clearly identifiable in your logs and on the dashboard.

Configure at init:

egisai.init(..., on_block="stub")

---

Configuration reference

Initialization parameters

Parameter	Default	Description
api_key	—	Required unless EGISAI_API_KEY is set. Your EgisAI SDK key (egis_live_…).
app	"default"	Logical application name; appears as an Agent in the dashboard for attribution.
env	"production"	Environment label (for example staging, prod). Free-form string for your own segmentation.
base_url	Hosted control plane	Override only when directed by EgisAI (for example dedicated regions or enterprise deployments).
on_block	"raise"	"raise" or "stub" — see above.
refresh_interval_seconds	10	How often to poll for policy updates if live streaming is unavailable.
enable_sse	True	Subscribe to live policy and configuration updates when supported.
enable_http_fallback	True	Optional patching of httpx / requests for broader HTTP visibility where enabled.
quiet	False	Set True to suppress the one-line startup banner on stderr.

Environment variables

Variable	Purpose
EGISAI_API_KEY	SDK API key if not passed as api_key=.
EGISAI_BASE_URL	Control plane base URL override when supplied by EgisAI.

Treat API keys as secrets—use environment variables or a secrets manager, never commit them to source control.

---

Policies (operator concepts)

Organizations configure policies in the dashboard. Typical categories include:

Category	Purpose (high level)
PII & secrets	Detect and block or mask categories such as government identifiers, payment data, and credential-shaped strings before model calls.
Content patterns	Allow or deny prompts or outputs matching operator-defined patterns.
Models & size	Restrict which model names may be called or cap prompt size.
Intent	Block requests that match dangerous or out-of-scope intent even when phrased obliquely or in another language.
Tools & connectors	Restrict tool, shell, or integration use when the model returns structured tool or command requests.

Exact rule JSON and ordering are managed in the product; the SDK consumes the published configuration and does not require you to embed policy documents in your repository.

---

Advanced: explicit context (optional)

For multi-tenant or test scenarios, you may override auto-detected context (for example agent identity) with egisai.set_context(**kwargs) as described in the package API. This is optional—the default path fingerprints agents from your application’s behavior.

---

Performance and availability

• Steady-state overhead is designed to stay on the order of a fraction of a millisecond for policy lookup per call after initialization and cache warm-up.
• Control plane connectivity — If the SDK cannot reach EgisAI at startup, your process can still run; policy enforcement may be limited until a successful connection and policy fetch. PII and other local checks remain in force where the engine can evaluate them. For your specific deployment’s behavior, refer to your contract and SECURITY.md.
• Audit delivery is asynchronous so network latency to EgisAI does not sit on the critical path of every model call.

---

Privacy and security

• Do not embed secrets in repository copies of this README.
• For vulnerability reporting, see SECURITY.md — please use the disclosed channel rather than public issues for security-sensitive matters.

A short summary suitable for architecture reviews:

• Governance evaluates prompts with respect to your organization’s policies before upstream invocation where applicable.
• Sensitive-content handling is architected so that raw regulated values are not sent to third-party LLMs as part of policy enforcement workflows described here.

---

Troubleshooting

Symptom	Things to check
RuntimeError: egisai.init() requires api_key	Set api_key= or EGISAI_API_KEY.
Policies never update	Network egress to your configured control plane; SSE disabled behind strict firewalls—polling still applies on an interval.
Calls succeed but dashboard stays empty	Confirm the SDK key matches the org you expect; verify process can reach the control plane for logging.
Blocked call raises unexpectedly	Review active policies in the dashboard; set on_block="stub" if you need non-throwing behavior.

---

Supported Python libraries

Library	Notes
openai ≥ 1.40	Chat Completions, Responses API, streaming where supported by the adapter.
anthropic ≥ 0.40	Messages API, streaming.
google-genai ≥ 1.0	client.models.generate_content, async, streaming.
google-generativeai ≥ 0.8	GenerativeModel.generate_content, streaming. Install via egisai[google-legacy].
httpx / requests	Optional broad HTTP capture when the fallback is enabled.

Minimum versions are guidance; pin in your own requirements.txt for reproducible builds.

---

Enforcement matrix

Different frameworks expose tool execution at different boundaries — some run the agentic loop in Python (where we can sit between the model and the tool), and a couple run it in a subprocess or on managed infrastructure. The matrix below is the honest, locked contract for what egisai can stop before it happens versus what it can only audit after the fact, per framework.

Two enforcement seams matter for SOC 2 / ISO 27001 / GDPR / HIPAA:

1. Tool / MCP call enforcement — block dangerous tool dispatches (deny_tool_call, deny_mcp_call, semantic_guard) before the tool runs.
2. Tool result enforcement — block / mask PII (pii_scan, deny_output_regex) in the tool's response before the model is shown it.

A row says enforced when the SDK can physically prevent the failure mode in question. It says advisory when the SDK observes after the fact and records the violation in the audit log but couldn't intervene.

Framework	Tier	Tool / MCP block	Tool result PII block	How it works
OpenAI	1	enforced	enforced (next call)	Output policy raises PermissionError before the response (with tool_calls) returns. Tool results round-trip Python; the next call's input phase scans them.
Anthropic	1	enforced	enforced (next call)	Output policy on Messages.create response with tool_use blocks. Tool result blocks in the next call's messages are scanned by the input phase.
Google GenAI (Gemini)	1	enforced	enforced (next call)	generate_content response with function_call parts is policy-gated before return; tool responses scanned on next call.
Google (legacy)	1	enforced	enforced (next call)	Same gate as Google GenAI; legacy google-generativeai package.
AWS Bedrock Converse	1	enforced	enforced (next call)	Converse / ConverseStream response gated; toolUse blocks blocked before caller dispatches; tool results scanned on next call.
HTTP fallback (httpx/requests)	1	enforced	enforced (next call)	Best-effort body parsing for unknown providers; the next request's payload text gets the same input-phase scan.
LangChain	1	enforced	enforced (next call)	Cascades to the underlying provider patch (OpenAI / Anthropic / Google).
OpenAI Agents	2	enforced	enforced (next call)	Identity-only wrap on Runner.run; cascades to inner OpenAI patch + input-phase scan on tool results.
CrewAI	2	enforced	enforced (next call)	Identity-only wrap on Agent.execute_task; cascades.
AutoGen	2	enforced	enforced (next call)	Identity-only wrap on AssistantAgent.run; cascades.
LangGraph	2	enforced	enforced (next call)	Identity-only wrap on Pregel.invoke; cascades.
LlamaIndex	2	enforced	enforced (next call)	Identity-only wrap on FunctionAgent.run; cascades.
Agno	2	enforced	enforced (next call)	Identity-only wrap on Agent.run / Agent.arun; cascades.
smolagents	2	enforced	enforced (next call)	Identity-only wrap on agent entry; cascades.
Strands Agents	2	enforced	enforced (next call)	Identity-only wrap on Agent.__call__; cascades.
Pydantic AI	2	enforced	enforced (next call)	Identity-only wrap on Agent.run; cascades.
Google ADK	2	enforced	enforced (next call)	Identity-only wrap on the ADK entry; cascades.
Claude Agent SDK	3a	enforced (PreToolUse)	enforced (PostToolUse)	Subprocess agent loop. We inject PreToolUse AND PostToolUse hooks. PreToolUse gates the dispatch; PostToolUse evaluates the tool's response and substitutes via updatedToolOutput / updatedMCPToolOutput before Claude is shown the result. On older SDK versions without the hooks field, falls back to advisory mode and the audit row is honestly labelled.
AWS Bedrock Agents	3b	advisory	advisory	Action Groups execute on AWS-managed infrastructure with no equivalent of PostToolUse. The patch records what happened but cannot prevent the tool dispatch OR substitute its result before the model sees it. Use the standalone bedrock-runtime Converse API or claude_agent_sdk for SOC 2 / GDPR-grade enforcement.

What you can rely on for every row above except Bedrock Agents:

• A deny_tool_call / deny_mcp_call / semantic_guard verdict on a tool call physically stops the tool from running.
• A pii_scan / deny_output_regex / semantic_guard verdict on a tool result either masks the result in place (action="sanitize") or refuses it (action="block") before the model is shown it. For Tier 1 + Tier 2 frameworks this happens on the next round trip's input phase; for the Claude Agent SDK it happens at the PostToolUse hook so the model never even sees the raw bytes for one turn.
• Input-side policies (PII scan, deny_regex, deny_model, max_prompt_chars, semantic_guard on the prompt) always run before the model is called.
• Sanitization rewrites the prompt locally before it reaches the provider.
• Aggregated OUTPUT replay — For claude_agent_sdk, OUTPUT policies that evaluate after MCP/tool payloads have been replayed from the CLI stamp verdict=block as enforcement_status="advisory" on the audit row (text-only breaches with hooks wired remain enforced). The withhold at your code boundary via on_block="raise" is unchanged — the distinction is for SOC 2 evidence about subprocess timing.
• Audit rows distinguish enforcement_status="enforced" (we actually prevented the action) from enforcement_status="advisory" (we observed after the fact or replayed MCP payloads before the aggregated OUTPUT evaluator ran). SOC 2 / GDPR auditors can query both states.

For Bedrock Agents specifically, see SECURITY.md — the limitation is publicly documented so customers can risk-assess accordingly.

---

Verifying PyPI artifacts (optional)

Releases are published to PyPI via automation. To verify a wheel cryptographically when verifying identity bindings published by the project:

pip download egisai==0.10.0 --no-deps
python -m sigstore verify identity \
  --cert-identity-regexp "https://github.com/EgisLabs/egisai-sdk/.+" \
  --cert-oidc-issuer "https://token.actions.githubusercontent.com" \
  egisai-0.10.0-py3-none-any.whl

Adjust the version to match the release you installed. A CycloneDX SBOM is attached to GitHub releases for supply-chain review.

---

Resources

Resource	URL
Website	egisai.co
Documentation	docs.egisai.co
Dashboard	app.egisai.co
PyPI	pypi.org/project/egisai
Repository & issues	github.com/EgisLabs/egisai-sdk
Changelog	CHANGELOG.md on GitHub
Security	SECURITY.md on GitHub

---

Licence

Apache License 2.0 — see the LICENSE file in the source repository.

---

EgisAI — runtime governance for AI agents.