llms.txt
# Traccia
> Traccia is an OpenTelemetry-native AI agent control plane: one layer to observe, evaluate, govern, and prove what production AI agents do. Real-time policies run inline through the `@govern()` SDK wrapper or the Traccia Gateway and can block, reshape, or hold an LLM or tool call for human approval before it runs. Traccia also covers tracing and cost attribution across 2,500+ models, prompt management and evaluation (datasets, scorers, experiments), and compliance evidence for the EU AI Act and HIPAA. Open-source Python and TypeScript SDKs integrate with LangChain, CrewAI, the OpenAI Agents SDK, Claude Code, and Cursor.
```bash
pip install traccia
```
Sign up or log in at the Traccia app (linked from the site header) to use the hosted dashboard.
## Product
- [Home](https://traccia.ai/): The AI agent control plane: observe, evaluate, govern, and prove.
- [AI Agent Governance](https://traccia.ai/governance/): Turn AI agent telemetry into governance and compliance evidence: policies, risk controls, audit trails, human oversight, EU AI Act and HIPAA support.
- [Pricing](https://traccia.ai/pricing/): Free, Team ($299/mo), Business ($799/mo), and Enterprise plans. Plans differ by events, retention, and active policies.
- [Security](https://traccia.ai/security/): Security practices and data handling.
- [Trust center](https://traccia.ai/trust-center/): Vendor trust, subprocessors, and certification roadmap.
- [Privacy](https://traccia.ai/privacy/): Privacy policy for Traccia and Algen AI Private Limited.
- [Terms](https://traccia.ai/terms/): Terms of service for the Traccia product.
- [Contact](https://traccia.ai/contact/): Support, partnerships, and company contact.
- [About](https://traccia.ai/about/): Company and mission.
## Blog
- [Blog](https://traccia.ai/blog/): Learn more about how Traccia is building the future of AI observability, policy enforcement, and governance.
- [Top 5 AI Agent Control Planes in 2026](https://traccia.ai/blog/top-ai-agent-control-planes/): Traccia, Lyzr Opencontroller, Microsoft Agent 365, Amazon Bedrock AgentCore, and ServiceNow AI Control Tower compared on identity, observability, evaluation, governance, audit, and vendor neutrality.
- [What Is AI Agent Observability?](https://traccia.ai/blog/ai-agent-observability/): The signals to capture, why logs and APM miss agent failures, and how Traccia's OpenTelemetry SDK traces every model call, tool, and cost.
- [What Is an AI Agent Control Plane?](https://traccia.ai/blog/what-is-an-ai-agent-control-plane/): The definition, the reference architecture, the 10 building blocks, and the one test that separates control from logging, in 8 diagrams.
- [How to Trace OpenAI Agents with Traccia](https://traccia.ai/blog/how-to-trace-openai-agents/): Complete trace visibility into every OpenAI Agents SDK LLM call, tool invocation, and turn boundary with token counts and USD costs.
- [How to Trace Gemini Agents with Traccia](https://traccia.ai/blog/how-to-trace-gemini-agents/): Gain instant visibility into every Gemini LLM call including prompts, completions, token usage, and cost with auto-instrumentation.
- [AI Agent Deleted the Production Database at 2 AM](https://traccia.ai/blog/ai-agent-deleted-production-database/): A hands-on guide to building governance guardrails before your autonomous agents cause an incident you can't undo.
- [Runtime Policy Enforcement in AI Agents](https://traccia.ai/blog/runtime-policy-enforcement-ai-agents/): Why AI agents need runtime policy enforcement, not just design-time guardrails: real incidents, the runtime decision flow, policy examples, the policy lifecycle, and how Traccia enforces policy in production.
- [Establishing Control Boundaries for AI Agents](https://traccia.ai/blog/establishing-control-boundaries-ai-agents/): The difference between system prompt instructions and agent control plane enforcement.
- [AI Agent Guardrails vs Policies](https://traccia.ai/blog/ai-agent-guardrails-vs-policies/): Guardrails shape judgement; policies constrain authority. Intent narrows authority, never grants it.
- [AI Agent Evaluation: From Transcripts to Execution-Based Evidence](https://traccia.ai/blog/agent-evaluation-playbook/): A research-led playbook for evaluating AI agents across outcomes, trajectories, reliability, safety, cost, and production behavior.
- [LLM Evaluation: Why It's the Foundation of Trustworthy AI Agents](https://traccia.ai/blog/llm-evaluation-foundation-trustworthy-ai-agents/): How to measure final-answer quality, tool selection, arguments, trajectories, groundedness, safety, and policy compliance.
- [Using Jev for AI Evals in Traccia](https://traccia.ai/blog/using-jev-for-ai-evals-in-traccia/): Practical guide to Jev Decision in Traccia Evals, from scorer setup to dataset experiments and SDK usage.
- [AI Agent Cost Attribution](https://traccia.ai/blog/ai-agent-cost-attribution/): AI agent cost attribution explained: why token dashboards fail, how to attribute AI spend by agent, user, department, workflow, and business unit, and why AI governance and runtime enforcement matter as much as observability.
- [Rethinking FinOps for Autonomous AI Agents](https://traccia.ai/blog/rethinking-finops-for-autonomous-ai-agents/): Traditional cloud FinOps was built for stable infrastructure. Autonomous agents demand a new discipline to track, attribute, and control dynamic costs.
- [Guardrails and Policy Enforcement for OpenAI Agents](https://traccia.ai/blog/openai-agents-guardrails-policy-enforcement/): How Traccia captures OpenAI Agents SDK guardrails and enforces policy with @govern.
- [What Is AI Agent Governance?](https://traccia.ai/blog/why-ai-agent-governance-became-inevitable/): Definitive guide to the four planes of control and standards mapping.
- [AI Observability vs AI Governance](https://traccia.ai/blog/ai-observability-vs-ai-governance/): The fail-closed test, EchoLeak case study, and latency gap.
- [Traccia vs Langfuse](https://traccia.ai/blog/langfuse-alternative/): Langfuse alternative — enforce, not just observe, with Visibility, Intelligence, Control, and Certification.
- [Traccia vs Arize AI](https://traccia.ai/blog/arize-alternative/): Arize AI / Phoenix alternative focused on agent-boundary policy enforcement and EU evidence from traces.
- [Traccia vs Azure AI Foundry](https://traccia.ai/blog/azure-ai-foundry-alternative/): Azure AI Foundry alternative — runtime agent enforce-and-prove on OpenTelemetry vs Azure-native model lifecycle and cloud security filters.
- [Traccia vs Credo AI](https://traccia.ai/blog/credo-ai-alternative/): Credo AI alternative — runtime OpenTelemetry control plane vs enterprise GRC.
- [Traccia vs Holistic AI](https://traccia.ai/blog/holistic-ai-alternative/): Holistic AI alternative — agent runtime enforce-and-prove vs portfolio GRC and bias audits.
- [Traccia vs Datadog](https://traccia.ai/blog/datadog-alternative/): Datadog alternative — runtime OpenTelemetry control plane vs LLM Observability / Agent Console.
- [Traccia vs Braintrust](https://traccia.ai/blog/braintrust-alternative/): Braintrust alternative — agent-boundary enforcement and EU evidence vs eval-first quality gates.
- [Traccia vs LangSmith](https://traccia.ai/blog/langsmith-alternative/): LangSmith alternative — framework-agnostic control plane vs LangChain-native tracing and evals.
- [Traccia vs Maxim](https://traccia.ai/blog/maxim-alternative/): Maxim alternative — enforce-and-prove on OpenTelemetry vs simulate, evaluate, and observe.
## Documentation
- [Docs overview](https://traccia.ai/docs/): Entry point for the AI agent control plane: visibility, intelligence, control, and certification across the SDK and platform.
- [SDK quickstart](https://traccia.ai/docs/sdk/quickstart/): Install and send your first trace in minutes.
- [Platform getting started](https://traccia.ai/docs/platform/getting-started/): Set up the hosted dashboard and workspace.
- [Platform OTLP ingestion](https://traccia.ai/docs/platform/ingestion/): HTTP and gRPC endpoints for SDK and native OTel clients.
- [FAQ](https://traccia.ai/docs/faq/): Common questions about Traccia.
- [Release Notes](https://traccia.ai/docs/release-notes/): Month-by-month platform release notes.
- [Redeem a coupon code](https://traccia.ai/docs/coupon/): Log in, open Billing & Plans, and apply a coupon code to unlock a paid plan for free.
## Documentation sections
- [SDK](https://traccia.ai/docs/sdk/): Index of SDK docs — decorators, prompts, evaluate(), exporters, metrics, and cost tracking.
- [Platform](https://traccia.ai/docs/platform/): Index of hosted platform docs — ingestion, dashboards, guardrails, and costs.
- [Concepts](https://traccia.ai/docs/concepts/): Index of core concepts — traces, spans, agents, and OpenTelemetry.
- [Guides](https://traccia.ai/docs/guides/): Index of guides and recipes for tool guards, tracing, experiments, multi-agent, and troubleshooting.
- [Reference](https://traccia.ai/docs/reference/): Index of API, CLI, configuration, event, and error references.
- [Integrations](https://traccia.ai/docs/integrations/): Index of framework and coding agent integrations: LangChain, CrewAI, OpenAI Agents, Claude Code, Cursor, and GitHub Copilot.
- [Compliance](https://traccia.ai/docs/compliance/): Index of compliance docs, including the EU AI Act and HIPAA.
## SDK
- [Installation](https://traccia.ai/docs/sdk/installation/): Python package setup and requirements.
- [Agent identity](https://traccia.ai/docs/sdk/agent-identity/): Name and version agents in traces.
- [@observe() decorator](https://traccia.ai/docs/sdk/observe-decorator/): Instrument functions and agents.
- [Auto-instrumentation](https://traccia.ai/docs/sdk/auto-instrumentation/): Automatic tracing for supported frameworks.
- [Exporters](https://traccia.ai/docs/sdk/exporters/): Send traces to Traccia, Jaeger, Grafana, and more.
- [Metrics](https://traccia.ai/docs/sdk/metrics/): Custom metrics alongside traces.
- [Governance in the SDK](https://traccia.ai/docs/sdk/governance/): observe vs govern, Preventive this-call stop, Detective next-wrap, and trace evidence.
- [Guardrail detection](https://traccia.ai/docs/sdk/guardrails/): Detect policy violations in agent runs.
- [Pricing and costs](https://traccia.ai/docs/sdk/pricing/): Token and cost attribution in traces.
- [Prompts](https://traccia.ai/docs/sdk/prompts/): Load versioned prompts with cache, fallback, and span identity.
- [Evaluate](https://traccia.ai/docs/sdk/evaluate/): Run a task and scorers over a dataset from Python or TypeScript; persist an experiment by default.
- [Advanced](https://traccia.ai/docs/sdk/advanced/): Power-user SDK patterns.
## Integrations
- [LangChain](https://traccia.ai/docs/integrations/langchain/): Trace LangChain chains and agents.
- [Claude Code](https://traccia.ai/docs/integrations/claude-code/): Native observability for Claude Code — sessions, tool traces, and cost labels with zero code changes.
- [Cursor](https://traccia.ai/docs/integrations/cursor/): Trace Cursor with one command. Paste an ingestion key, restart Cursor, then open Sessions.
- [GitHub Copilot](https://traccia.ai/docs/integrations/github-copilot/): Trace GitHub Copilot in VS Code and the Copilot CLI through its built-in OpenTelemetry export. Agent turns, LLM and tool spans, sessions, and estimated cost, with no extension or SDK.
- [CrewAI](https://traccia.ai/docs/integrations/crewai/): Observe CrewAI crews and tasks.
- [OpenAI Agents SDK](https://traccia.ai/docs/integrations/openai-agents/): Instrument OpenAI Agents workflows.
- [Gemini](https://traccia.ai/docs/integrations/gemini/): Traccia automatically instruments Google's Gemini SDK for full observability of LLM calls.
## Platform
- [OTLP ingestion](https://traccia.ai/docs/platform/ingestion/): Platform HTTP/gRPC endpoints, auth, and client-specific config.
- [Dashboard](https://traccia.ai/docs/platform/dashboard/): Workspace overview and navigation.
- [Traces](https://traccia.ai/docs/platform/traces/): Inspect runs, spans, and timelines.
- [Guardrail posture](https://traccia.ai/docs/platform/guardrails/): Monitor guardrail health across agents.
- [Policies](https://traccia.ai/docs/platform/policies/): Preventive vs Detective, the full template list, Save Draft or Activate Policy, Observe / Warn / Block, Decision Log, violations, email alerts, and ingest never rejects a trace.
- [Spend Cap](https://traccia.ai/docs/platform/policies/spend-cap/): Daily USD budget for model calls, checked before the provider is charged. SDK and Gateway.
- [Model Boundary](https://traccia.ai/docs/platform/policies/model-boundary/): Allow only the models you list, with an optional fallback model. SDK and Gateway.
- [Loop Cap](https://traccia.ai/docs/platform/policies/loop-cap/): Maximum tool calls in one run. SDK only.
- [Refund Guard](https://traccia.ai/docs/platform/policies/refund-guard/): Deny refunds above a cutoff, with optional human approval for a middle band. SDK only.
- [Purchase Guard](https://traccia.ai/docs/platform/policies/purchase-guard/): Deny orders and buys above a cutoff, with optional human approval. SDK only.
- [High-Risk Tool](https://traccia.ai/docs/platform/policies/high-risk-tool/): Deny a tool by name before it runs. SDK only.
- [Dangerous Shell Commands](https://traccia.ai/docs/platform/policies/dangerous-shell-commands/): Deny a shell tool when its command contains text you listed. Not a sandbox. SDK only.
- [Freshness Guard](https://traccia.ai/docs/platform/policies/freshness-guard/): Deny a write when the last read is older than your limit. SDK only.
- [Retrieval Completeness](https://traccia.ai/docs/platform/policies/retrieval-completeness/): Deny an answer when no retrieval happened or too few chunks came back. SDK only.
- [Unique Customer Cap](https://traccia.ai/docs/platform/policies/unique-customer-cap/): Limit how many distinct customers an agent can open in a window. SDK only.
- [Prompt Pin](https://traccia.ai/docs/platform/policies/prompt-pin/): Deny a model call when a Traccia prompt was loaded from the wrong label or version. SDK only.
- [Cost Spike Alert](https://traccia.ai/docs/platform/policies/cost-spike-alert/): After Run alert when total spend over an hour, day, or week passes a limit. Opens a violation after traces land.
- [Slow Run Alert](https://traccia.ai/docs/platform/policies/slow-run-alert/): After Run alert when run duration (P95, average, or slowest) passes a limit.
- [Tool Storm Alert](https://traccia.ai/docs/platform/policies/tool-storm-alert/): After Run alert when a single run makes too many tool calls. Use Loop Cap to stop a run live.
- [Unapproved Model Alert](https://traccia.ai/docs/platform/policies/unapproved-model-alert/): After Run alert when a run uses a model outside your list. Use Model Boundary to stop the call.
- [Output Token Limit](https://traccia.ai/docs/platform/policies/output-token-limit/): After Run alert when response length passes a token budget.
- [Input Token Limit](https://traccia.ai/docs/platform/policies/input-token-limit/): After Run alert when prompts grow past a token budget.
- [Gateway](https://traccia.ai/docs/platform/gateway/): Policy enforcement for OpenAI, Anthropic, or Gemini without the SDK. Not a tracing product. Gateway-only traces are one LLM span, nowhere close to an SDK timeline. Use SDK only, Gateway only, or SDK + Gateway for traffic the SDK cannot wrap.
- [Gateway OpenAI SDK](https://traccia.ai/docs/platform/gateway/openai/): Official OpenAI client. base_url / baseURL is https://gateway.traccia.ai/openai/v1. Chat Completions only.
- [Gateway Anthropic SDK](https://traccia.ai/docs/platform/gateway/anthropic/): Official Anthropic client. base_url is https://gateway.traccia.ai/anthropic with no /v1. Messages API only.
- [Gateway Gemini SDK](https://traccia.ai/docs/platform/gateway/gemini/): google-genai http_options.base_url / httpOptions.baseUrl is https://gateway.traccia.ai/google. Developer API only, not Vertex.
- [Gateway LangChain](https://traccia.ai/docs/platform/gateway/langchain/): ChatOpenAI and ChatAnthropic. Force Chat Completions. Responses API is not this Gateway.
- [Gateway LiteLLM](https://traccia.ai/docs/platform/gateway/litellm/): completion() with api_base and extra_headers. Prefix openai/ or anthropic/.
- [Gateway Vercel AI SDK](https://traccia.ai/docs/platform/gateway/vercel-ai-sdk/): createOpenAI must use openai.chat(). Anthropic baseURL must include /v1.
- [Gateway LlamaIndex](https://traccia.ai/docs/platform/gateway/llamaindex/): OpenAI(api_base, default_headers). Not OpenAIResponses.
- [Gateway CrewAI](https://traccia.ai/docs/platform/gateway/crewai/): LLM(base_url, extra_headers, is_litellm=True). LiteLLM under the hood.
- [Gateway OpenAI Agents SDK](https://traccia.ai/docs/platform/gateway/openai-agents/): AsyncOpenAI plus set_default_openai_api("chat_completions").
- [Gateway n8n](https://traccia.ai/docs/platform/gateway/n8n/): HTTP Request with both Traccia headers. Stock OpenAI credential currently exposes one extra header.
- [Gateway HTTP](https://traccia.ai/docs/platform/gateway/http/): cURL and any HTTP client. Full path plus both Traccia headers.
- [Prompts](https://traccia.ai/docs/platform/prompts/): Versioned prompt library; Metrics tab (By Version, Recent Calls); Connections for agents and AI systems; promote to production with experiment evidence.
- [Prompt Playground](https://traccia.ai/docs/platform/prompt-playground/): Compare prompts and models side by side with latency, tokens, and cost before you promote.
- [Datasets](https://traccia.ai/docs/platform/datasets/): Curate test cases (CSV, manual, or from production traces), edit name/description, optional JSON Schema, and scorers for Prompt Playground experiments.
- [Scorers](https://traccia.ai/docs/platform/scorers/): Built-in scorers, LLM-as-judge, and restricted custom code for dataset experiments.
- [Experiments](https://traccia.ai/docs/platform/experiments/): Saved graded runs, runs-in-progress strip, baseline vs candidate compare (including multi-panel pairing), and evidence to attach when promoting a prompt version.
- [Compliance Hub](https://traccia.ai/docs/platform/governance/): Org-wide AI system registry, human review queue, incidents, audit bundles, and admin activity log.
- [EU AI Act](https://traccia.ai/docs/compliance/eu-ai-act/): Opt-in module — AI system registry, human review, FRIA drafts, and labeled audit exports for EU teams.
- [HIPAA](https://traccia.ai/docs/compliance/hipaa/): Opt-in module — PHI-aware registry fields, safeguard checklists, soft SDK warnings, and labeled evidence exports (no signed Traccia BAA).
- [Costs](https://traccia.ai/docs/platform/costs/): Cost analytics and budgets.
- [Teams](https://traccia.ai/docs/platform/teams/): Collaboration and access control.
## Concepts
- [Traces](https://traccia.ai/docs/concepts/trace/): Root spans and trace structure.
- [Spans](https://traccia.ai/docs/concepts/span/): Nested operations within a trace.
- [Agents](https://traccia.ai/docs/concepts/agents/): How Traccia models agents.
- [OpenTelemetry](https://traccia.ai/docs/concepts/opentelemetry/): OTel compatibility and semantics.
## Guides
- [Cap refunds and purchases](https://traccia.ai/docs/guides/cap-refunds-and-purchases/): Stop an agent from refunding or buying above a cutoff in the policy currency, before that function runs. SDK only. Not Spend Cap. Not Stripe.
- [Approve a refund or purchase](https://traccia.ai/docs/guides/approve-a-refund-or-purchase/): Let a person approve the middle band of refunds and purchases an AI agent may not make alone. Set the amounts, review requests in the Approvals inbox, and run the approved action from your own server with a signed webhook. SDK only.
- [Stop a dangerous tool](https://traccia.ai/docs/guides/stop-a-dangerous-tool/): Deny a named function, or a shell command that contains text you listed. Not a sandbox. SDK only.
- [Local Jaeger setup](https://traccia.ai/docs/guides/local-jaeger/): Run Jaeger locally
