TealTiger

AI Agent Security & Governance SDK

Deterministic governance, guardrails, cost tracking, and policy management for LLM applications. Open source. TypeScript + Python. Works with any provider.

Website · Documentation · Examples · Discord · Contributing

---

⚡ 60-second quickstart

Install: npm install tealtiger or pip install tealtiger, then wrap one existing OpenAI call:

import { TealOpenAI } from 'tealtiger';
const client = new TealOpenAI({ apiKey: process.env.OPENAI_API_KEY, guardrails: { promptInjection: true } });
const res = await client.chat.completions.create({ model: 'gpt-4o-mini', messages: [{ role: 'user', content: 'Hello!' }] });
console.log(res.security?.decision ?? 'ALLOW');

import os
from tealtiger import TealOpenAI
client = TealOpenAI(api_key=os.environ["OPENAI_API_KEY"], guardrails={"prompt_injection": True})
print(client.chat.completions.create(model="gpt-4o-mini", messages=[{"role": "user", "content": "Hello!"}]).security.decision)

ALLOW
Governance receipt emitted; cost and guardrails tracked.

Next: full Quick Start and examples.

---

🔭 observe() — Zero-Config Instrumentation (v1.4)

One line adds cost tracking, audit logging, PII detection, and behavioral baselines to any LLM client. No config files, no policy definitions.

import { observe, freeze } from 'tealtiger';
const client = observe(new OpenAI());  // done — all calls are now instrumented
console.log(client.getCost());         // { totalCost: 0.0023, requestCount: 1, ... }

from tealtiger.observe import observe, freeze
client = observe(OpenAI())             # done — all calls are now instrumented
print(client.get_cost())               # ObserveCostSummary(total_cost=0.0023, ...)

What you get automatically: per-request cost tracking across 12 providers, structured audit log with correlation IDs, behavioral baseline (P50/P95/P99), PII detection in REPORT_ONLY mode, and an instant kill switch via freeze(). Under 5ms overhead per call.

See examples/observe-quickstart.ts and examples/observe_quickstart.py.

---

What is TealTiger?

TealTiger is an open-source SDK that provides deterministic governance for AI agents. It enforces security policies, tracks costs, and produces structured evidence — all at runtime, with no infrastructure required.

Looking for the source code? This is the hub repo. The SDK source lives in the language-specific repos:

• TypeScript SDK: tealtiger-typescript-prod
• Python SDK: tealtiger-python-prod

Or clone this repo with submodules: git clone --recurse-submodules https://github.com/agentguard-ai/tealtiger.git

Unlike probabilistic safety filters, TealTiger uses deterministic policy evaluation: same input + same policy = same decision, every time. Every governance verdict is reconstructable, traceable to the human who authored the policy, and exportable as structured evidence (SARIF, JUnit XML, JSON).

Key principle: Governance should be an engineering property embedded in the runtime — not a document reviewed after the fact.

---

🚀 Quick Start

TypeScript

npm install tealtiger

import { TealOpenAI } from 'tealtiger';

const client = new TealOpenAI({
  apiKey: process.env.OPENAI_API_KEY,
  guardrails: {
    piiDetection: true,
    promptInjection: true,
    contentModeration: true,
  },
  budget: {
    maxCostPerRequest: 0.50,
    maxCostPerDay: 10.00,
  },
});

const response = await client.chat.completions.create({
  model: 'gpt-4',
  messages: [{ role: 'user', content: 'Hello!' }],
});
// Guardrails enforced. Cost tracked. Evidence produced.

Python

pip install tealtiger

from tealtiger import TealOpenAI

client = TealOpenAI(
    api_key=os.getenv("OPENAI_API_KEY"),
    guardrails={
        "pii_detection": True,
        "prompt_injection": True,
        "content_moderation": True,
    },
    budget={
        "max_cost_per_request": 0.50,
        "max_cost_per_day": 10.00,
    },
)

response = client.chat.completions.create(
    model="gpt-4",
    messages=[{"role": "user", "content": "Hello!"}],
)
# Guardrails enforced. Cost tracked. Evidence produced.

---

✨ Features

🛡️ Security Guardrails

• PII Detection — Detect and redact sensitive information automatically
• Prompt Injection Prevention — Block malicious prompt injection attempts
• Content Moderation — Filter toxic, harmful, or inappropriate content
• Secret Detection — 500+ patterns across 9 categories with confidence scoring
• Custom Rules — Define your own security policies

💰 Cost Governance

• Budget Enforcement — Hard limits per request, session, and day
• Cost Tracking — Real-time monitoring across all providers
• Cost Alerts — Notifications at configurable thresholds
• Circuit Breakers — Prevent runaway cost loops automatically

🔌 12 LLM Providers

• OpenAI — GPT-4, GPT-4o, GPT-3.5
• Anthropic — Claude 3.5, Claude 3
• Google Gemini — Multimodal support
• AWS Bedrock — Claude, Titan, Jurassic, Command, Llama
• Azure OpenAI — Deployment-based routing
• Cohere — Chat, RAG, embeddings
• Mistral AI — European data residency
• DeepSeek — Cost-efficient reasoning models
• Groq — Ultra-low latency inference
• Together AI — Open-source model hosting
• HuggingFace TGI — Self-hosted inference
• xAI (Grok) — Real-time knowledge

🔌 Platform Adapters

• AWS Bedrock Agents — Native guardrail adapter
• AWS AgentCore — Pre/post action governance plugin
• Azure AI Agent Service — Tool-call pipeline middleware

🏗️ Governance Architecture

• Deterministic Policy Evaluation — No LLM in the governance path
• Structured Evidence — Every decision produces a reconstructable record
• Cryptographic Proof — Merkle trees + RFC 3161 timestamping (TealProof)
• Non-Human Identity (NHI) — Agent lifecycle, scope enforcement, Zero Standing Privilege
• FREEZE Rules — Immutable emergency kill switches with tamper detection
• Correlation IDs — End-to-end traceability across the decision chain
• Policy Traceability — Every verdict traces to the human policy author
• OWASP Agentic Top 10 — Zero-config policy pack covering all 10 ASI risks

---

🗺️ Governance Coverage

Dimension	What it does	Module
🛡️ Security	Secret detection (500+ patterns), prompt injection, PII, content moderation, Unicode normalization, encoded output detection	TealSecrets TealGuard
🔑 Identity	Non-Human Identity lifecycle, scope enforcement, Zero Standing Privilege, agent attestation	TealEngine (NHI)
⚡ Reliability	Circuit breakers, retry budgets, fallback chains, deterministic degradation	TealCircuit TealReliability
🧠 Memory	Write provenance, instruction injection detection, exfiltration prevention, scope enforcement	TealMemory
💰 Cost	Governance-owned ceilings, anomaly detection, reasoning-token budgets, per-agent attribution	TealMonitor
📋 Evidence	Cryptographic receipts (Merkle + RFC 3161), SARIF export, OTel spans, SIEM integration	TealProof TealAudit
⚙️ Policy	FREEZE rules, PLAN_ONLY mode, hot-swap bundles, anti-tamper, automation levels	TealEngine
🔄 Workflow	Declarative YAML governance workflows, org-level inheritance, floor enforcement	TealFlow
📊 Drift	Behavioral drift detection, statistical baselines, model output regression	TealDrift
⏱️ Temporal	Session TTL, cooldown periods, time-of-day restrictions	TealTemporal
🔍 Registry	MCP definition-drift monitoring, tool description scanning, adapter composition allowlist	TealRegistry
🧠 Classification	Local ONNX ML inference (≤20ms), ensemble modes, regex+ML combination	TealClassifier

Design principle: No LLM in the governance path. Same input + same policy = same decision, every time.

---

📦 SDKs

Language	Source Code	Package	Install
TypeScript	tealtiger-typescript-prod	npm	npm install tealtiger
Python	tealtiger-python-prod	PyPI	pip install tealtiger

🔌 Framework Adapters

Framework	Package	Install
LangChain	langchain-tealtiger	pip install langchain-tealtiger
Vercel AI SDK	tealtiger-ai-sdk	npm install tealtiger-ai-sdk
PydanticAI	pydanticai-tealtiger	pip install pydanticai-tealtiger
Haystack	haystack-tealtiger	pip install haystack-tealtiger
CAMEL-AI	camelai-tealtiger	pip install camelai-tealtiger

🔗 Infrastructure Integrations

Platform	What it provides	Install
Dakera	Persistent governance state backend (cost storage, decision receipts, delegation chains via KG)	pip install dakera[tealtiger]
AG2 Beta	Governance middleware Extension for AG2 Beta agents	pip install ag2-tealtiger
Portkey Gateway	Webhook guardrail for Portkey AI Gateway	Example
Daytona	Pre-execution governance for sandboxed code execution	Example

---

📚 Documentation

• Why your AI agent needs a budget — cost governance for LLM apps
• OWASP Agentic Top 10 — practical defenses with TealTiger
• Quick Start Guide
• Security Guardrails
• Cost Governance
• Provider Setup
• FAQ
• Why TealTiger?
• Cross-SDK Feature Parity Matrix
• Governance Event Store Indexes
• Error Code Reference
• Troubleshooting
• Contributing Guide
• Security Policy
• Code of Conduct
• Roadmap

Badge

Use the TealTiger badge to show that a project is governed by deterministic agent security and cost policies.

Light badge:

[![Governed by TealTiger](https://raw.githubusercontent.com/agentguard-ai/tealtiger/main/assets/badges/governed-by-tealtiger.svg)](https://github.com/agentguard-ai/tealtiger)

Dark badge:

[![Governed by TealTiger](https://raw.githubusercontent.com/agentguard-ai/tealtiger/main/assets/badges/governed-by-tealtiger-dark.svg)](https://github.com/agentguard-ai/tealtiger)

Python Hugging Face TGI Quickstart

Use examples/python/huggingface_tgi_quickstart.py to try the guarded Hugging Face Text Generation Inference provider from the Python SDK.

export HF_API_TOKEN="your-hugging-face-token"
export HF_TGI_ENDPOINT="https://your-endpoint.endpoints.huggingface.cloud"
export HF_TGI_MODEL="meta-llama/Meta-Llama-3.1-8B-Instruct"

python examples/python/huggingface_tgi_quickstart.py

The example enables guardrail and cost-tracking configuration, sends one sample chat request, then prints the response, token usage, estimated cost, provider, and correlation ID. Use placeholder values in docs and .env.example files; never commit a real HF_API_TOKEN.

TealEngine Policy Schema

Use schemas/tealtiger-policy.schema.json for editor autocomplete and validation when authoring TealEngine policy JSON or YAML files. JSON policy files can include:

{
  "$schema": "./schemas/tealtiger-policy.schema.json"
}

For YAML policies, configure your editor's YAML schema mapping to point policy files such as tealtiger-policy.yml at ./schemas/tealtiger-policy.schema.json.

Validate Policy Files

Use the policy validator script to check a TealTiger policy JSON file before using it in CI/CD or runtime governance:

npm install
npx ts-node scripts/validate-policy.ts ./my-policy.json

You can also run the npm script:

npm run validate:policy -- ./my-policy.json

The validator loads schemas/tealtiger-policy.schema.json, prints schema validation errors, exits 0 when the policy is valid, and exits 1 when the policy is invalid.

---

🐯 Build With Us — Early Contributor Program

TealTiger is open source and we're looking for early contributors to shape the future of AI agent governance.

What You Can Work On

Area	Examples	Difficulty
🔍 Secret Detection	New detection patterns, custom categories	🟢 Beginner
📝 Documentation	Guides, examples, API docs, typo fixes	🟢 Beginner
🧪 Tests	Unit tests, property-based tests, integration tests	🟡 Intermediate
🔌 Integrations	LangChain, CrewAI, AG2, LlamaIndex middleware	🟡 Intermediate
💾 Memory Adapters	Redis, Pinecone, Weaviate, ChromaDB adapters	🟡 Intermediate
🔄 CI/CD Templates	Jenkins, Azure Pipelines, Bitbucket Pipelines	🟡 Intermediate
🏗️ Core Modules	Governance engine, evidence export, policy evaluation	🔴 Advanced

What Early Contributors Get

• 🏆 Named in CONTRIBUTORS.md and release notes
• 🎖️ "Founding Contributor" badge — first 25 merged PRs get permanent recognition
• 📣 Shoutout on TealTiger social channels (LinkedIn, X, Dev.to)
• 🔑 Early access to upcoming governance features before public release
• 💬 Direct access to the core team via GitHub Discussions
• 📝 Co-authorship opportunity on technical blog posts

Get Started

# 1. Star this repo (it helps!)

# 2. Fork and clone the SDK you want to contribute to:
# TypeScript SDK:
git clone https://github.com/agentguard-ai/tealtiger-typescript-prod.git
# Python SDK:
git clone https://github.com/agentguard-ai/tealtiger-python-prod.git

# 3. Pick a "good first issue"
# https://github.com/agentguard-ai/tealtiger/issues?q=label%3A%22good+first+issue%22

# 4. Submit a PR
# 5. Join the team 🐯

See CONTRIBUTING.md for detailed guidelines.

---

🗺️ Roadmap

Current: v1.3.0 — Autonomous Agent Governance (Released May 18, 2026)

• TealEngine v1.3 with pre/post evaluation pipeline, FREEZE rules, automation levels
• Non-Human Identity (NHI) governance with Zero Standing Privilege
• TealProof — cryptographic governance receipts (Merkle + RFC 3161)
• TealFlow — declarative YAML governance workflows
• TealClassifier — local ONNX ML inference (≤20ms)
• TealDrift, TealState, TealTemporal — behavioral, context, and session governance
• TealMonitor v2 — governance-owned cost ceilings, anomaly detection
• OWASP Agentic Top 10 policy pack (zero-config)
• 12 LLM providers + 3 platform adapters (Bedrock, AgentCore, Azure)
• Full Python SDK parity

Next: v1.4.0 — Zero-Config Adoption & MCP Governance (July 2026)

• observe(client) — 1-line auto-instrumentation, zero config, instant visibility
• Local CLI dashboard (npx tealtiger dashboard)
• Progressive disclosure: observe → suggest → enforce → govern
• 8 framework adapters: LangChain, LangGraph, CrewAI, AG2, LlamaIndex, CAMEL-AI, Haystack, Vercel AI SDK
• MCP governance: tool validation, per-identity grants, argument-level policies
• Tool poisoning & rug pull defense
• Runaway loop detection & per-trace token budgets
• TEEC v2.1 Execution Receipts (cryptographic evidence)
• EU AI Act, NIST AI RMF, ISO 42001 compliance mappings
• Universal Adapter Governance Contract — enforced across all framework adapters:
  • Every proposed tool call, budget change, freeze/unfreeze, and delegated handoff creates a decision record before execution
  • Decision binds: agent_id, turn_id, parent_turn_id, action kind, tool name, canonical args digest, policy digest, decision source, expiry, delegation scope
  • Execution outcome backlinks to decision_id and effective args digest that actually ran
  • Denial, timeout, and require-approval become visible terminal or pending states (no silent transcript gaps)
  • Per-actor enforcement: channel/hub approval does not become ambient authority for every agent
  • Delegation creates scope-bound authorization: delegatee emits its own decision and outcome record
  • Invariant tests: same payload in two turns produces two decision IDs; approval for one agent does not authorize another; revised args create a new pending decision; timeout produces durable terminal result; retry returns prior terminal state instead of re-running side effect
• Supply Chain Integrity Verification — validate AI dependency hashes before agent execution (response to LiteLLM supply chain attack, March 2026)
• A2A Protocol Governance — per-task access scoping, token lifetime enforcement, consent flow governance for Agent2Agent interactions (A2A GA under Linux Foundation, 150+ orgs)
• MCP Tool Integrity Monitoring — definition-drift detection, tool description scanning, adapter composition allowlist, rug-pull defense with cryptographic pinning
• Kill Switch Latency SLA — guaranteed <5ms FREEZE propagation with published benchmarks, quantified time-to-halt for compliance
• Shadow Agent Detection — discover ungoverned agents in the environment, surface unmonitored tool calls and untracked cost
• AI Dependency SBOM — generate Software Bill of Materials for agent dependencies, model provenance, and MCP server inventory
• Dakera Integration (shipped) — persistent governance state backend for distributed deployments (pip install dakera[tealtiger]). Cost records, decision receipts, and delegation chains survive restarts via agent-scoped memory with importance-weighted retention.
• Pack Hunt Defense — multi-agent coordinated attack detection:
  • Cross-session intent correlation: link related sessions via session_group_id, query receipts across sessions by agent + time window
  • Reassembly detection: flag when N individually-benign requests from different agents/sessions converge on the same output scope within a time window (deterministic, no LLM)
  • Multi-model coordination detection: track which model produced which output, flag when output from model A becomes input to model B without governance evaluation in between

Planned: v1.5.0 — Enterprise Platform (Q3 2026)

• Multi-tenancy with complete data isolation
• RBAC (Owner, Admin, Policy Author, Viewer, Auditor)
• SSO via SAML 2.0 / OIDC (Okta, Azure AD, Google)
• SIEM export (Splunk, Elastic, Sentinel, Datadog)
• Policy staging, dry-run mode, canary deployments
• Scheduled compliance reports & executive dashboard

Future: v2.0.0 — SaaS Security Platform (Q1 2027)

• Full SaaS control plane (CSPM/CWPP model for AI agents)
• CISO executive console with governance health scoring
• TealTiger Operator & Agent for Kubernetes
• Shadow AI detection (discover ungoverned agents)
• Remote kill switch from SaaS console
• CloudEvents, OpenTelemetry, Backstage plugin

All features maintain: in-process <5ms, zero-config entry, no LLM in governance path, offline-capable.

---

🌟 Community

• Discord: Join TealTiger Community
• GitHub Discussions: Ask questions, share ideas
• LinkedIn: TealTiger
• X (Twitter): @TealtigerAI
• Documentation: docs.tealtiger.ai
• Blog: blogs.tealtiger.ai
• Playground: playground.tealtiger.ai
• Email: reachout@tealtiger.ai

---

🔒 Security

TealTiger is committed to responsible open-source security practices.

For vulnerability reports, see our Security Policy.

---

📄 License

TealTiger is Apache 2.0 licensed.

---

🙏 Acknowledgments

Built with ❤️ by the TealTiger team and contributors.

---

👥 Contributors

Want to contribute? Check out our CONTRIBUTING.md guide!

---

⭐ Star this repo if you believe AI agents need governance, not just guardrails.

Report Bug · Request Feature · Ask Question