Skip to main content
TealTiger is a runtime AI governance SDK. It provides deterministic policy enforcement for AI agents and LLM applications — covering security, cost, identity, evidence, and 10 governance domains. No infrastructure required. Governance runs inside your application boundary.
Guardrails focus on content filtering (input/output). TealTiger is a full governance framework:
  • 10 governance domains (not just content)
  • Deterministic — no LLM in the governance path
  • Separation of duties — governance teams define policy, developers consume
  • Cryptographic evidence — Merkle receipts, RFC 3161 timestamping
  • Non-human identity — agents as first-class principals with lifecycle
  • Anti-tamper — signed bundles, FREEZE rules, forbidden config keys
See Why TealTiger for the full comparison.
Yes. Apache 2.0 license. The SDK will always remain free and open source. There is no SaaS dependency — everything runs locally in your application.
  • TypeScript — npm install tealtiger
  • Python — pip install tealtiger
Both SDKs have full feature parity (12 providers, all v1.3 modules).
12 providers: OpenAI, Anthropic, Google Gemini, AWS Bedrock, Azure OpenAI, Cohere, Mistral, DeepSeek, Groq, Together AI, xAI (Grok), HuggingFace TGI. Plus platform adapters for AWS Bedrock Agents, AWS AgentCore, and Azure AI Agent Service.
No. TealTiger’s governance path is 100% deterministic — rule-based evaluation, pattern matching, and threshold checks. No LLM is involved in any decision. Same input + same policy = same decision, every time. TealClassifier uses a local ONNX model for content classification, but the final governance decision is always rule-based.
  • Full v1.3 pipeline: < 25ms p99
  • Basic evaluation (4 modules): < 15ms p99
  • PLAN_ONLY mode: < 10ms p99
  • Pre-evaluation (FREEZE + NHI): < 2ms
For context, LLM calls themselves take 500ms-5s. TealTiger adds negligible overhead.
No. In the v1.3 separation of duties model:
  • FREEZE rules cannot be modified or disabled by application code
  • Signed bundles reject tampering (Ed25519 verification)
  • Forbidden config keys (disable_enforcement, bypass, allow_all) are rejected
  • Governance-owned cost ceilings cannot be raised by application code
  • All bypass attempts are logged as TAMPER_ATTEMPT security events
See Governance at Scale.
  • OWASP Agentic Top 10 — Complete coverage (ASI-01 through ASI-10)
  • NIST AI RMF — Cross-mapped controls
  • ISO 42001 — AI management system alignment
  • EU AI Act — Relevant articles referenced
  • SOC 2 — Audit evidence and separation of duties
  • HIPAA — PHI governance in agent memory
See Compliance Mapping.
Then follow the Quickstart — governance running in 5 minutes. For enterprise deployments, see Governance at Scale.
Immutable safety controls that cannot be bypassed by any code path. They persist across policy hot-swaps, process restarts, and configuration changes. Use them for actions that must NEVER happen (e.g., database.drop in production). See FREEZE Rules.
TealTiger treats AI agents as first-class principals with identity, lifecycle (active/suspended/revoked), scope enforcement, and Zero Standing Privilege. Every agent must be registered, verified, and scoped — just like human users in IAM. See NHI Governance.
No agent holds permanent elevated permissions. Every access to a protected resource requires a Just-In-Time (JIT) grant that is time-bounded, scope-limited, and audited. Grants expire automatically. See NHI Governance.
Every governance decision produces a cryptographic receipt (SHA-256 hash appended to a Merkle tree). Periodically, Merkle roots are anchored to an RFC 3161 Time Stamping Authority. This creates tamper-evident proof that governance was enforced at a specific time. A standalone Verification SDK allows third-party audit without TealTiger installed. See TealProof.
Yes. TealTiger is framework-neutral. It works with any agent framework — LangChain, CrewAI, AutoGen, LlamaIndex, MCP, or direct SDK calls. See Integrations.
TealTiger can run as:
  • Embedded SDK — inside your application (recommended)
  • Sidecar container — HTTP API alongside your pod
  • Helm chart — for cluster-wide deployment
See Deployment.
No. TealTiger is SDK-only by design. There is no hosted service, no control plane, no external dependency. Governance runs entirely inside your application boundary. This is intentional — it means no data leaves your infrastructure for governance purposes.