Traccia

Traccia is an AI agent control plane for observing activity, evaluating outputs, setting governance controls, and enforcing runtime policies. Its Python and TypeScript SDK records runs, sessions, tool calls, prompts, and errors. Spend limits, model allowlists, and loop caps can stop a call; detective rules instead record issues after a run. Evaluation tools combine prompts, datasets, scorers, and experiments to assess outputs before promotion. Listed integrations include LangChain, Claude Code, CrewAI, OpenAI Agents SDK, and Gemini SDK. Traccia says it works with any LLM, with automatic token and cost tracking for OpenAI and Anthropic and OpenTelemetry tracing for HTTP-based providers. Its open-source SDK can export traces to compatible backends without an account. Optional pattern-based PII masking can run before spans leave a user's process; ML PII detection is planned. The free Hobby plan includes 50K events and seven days of retention. Paid plans start at $99 a month; Observe lists 500K events and 30 days of retention. Traccia says SOC 2 Type II certification is in progress and that it does not offer a signed Business Associate Agreement.

Traccia offers agent tracing, evaluation, and runtime controls, with a free plan and paid plans starting at $99 a month. Its stated HIPAA limitation and in-progress SOC 2 Type II certification are important considerations for organizations with compliance requirements.

Traccia suits teams that need to inspect agent runs, evaluate outputs, or enforce limits on calls. Teams handling health information should note that the product does not offer a signed Business Associate Agreement.

Traccia is an AI-agent control plane that combines tracing, evaluation, governance, and runtime policy enforcement. It fits teams building Python or TypeScript agents that need to see tool use and errors while also placing limits on execution. The free tier is useful for a small deployment, but organizations needing a signed HIPAA agreement should choose another product.

Overview

Traccia brings observability and intervention into one platform. Its SDK records agent runs, sessions, tool calls, prompts, and errors, giving developers a workflow-level view rather than a narrow log of model requests. Evaluation features cover prompts, datasets, scorers, and experiments, so teams can assess outputs before promoting changes.

The open-source SDK can send traces to Jaeger, Grafana Tempo, Zipkin, or any OpenTelemetry-compatible backend without a Traccia account. That is a practical route for teams that want local control or an existing telemetry stack, while hosted plans add retention, quotas, and managed controls.

Key features

Python and TypeScript tracing is supported, with documented integrations for LangChain, Claude Code, CrewAI, OpenAI Agents SDK, and Gemini SDK. Traccia says it works with any LLM. Automatic token and cost tracking is specified for OpenAI and Anthropic; HTTP-based providers can be traced through OpenTelemetry, so broader coverage may require that instrumentation path.

Runtime controls are more consequential than passive monitoring. Spend limits, model allowlists, and loop caps can stop a call while it is running. Detective rules record behavior after a run instead, making them useful for review but not prevention. Session replay, prompt and tool tracing, and cost tracking support incident analysis and budget oversight.

Optional pattern-based PII masking can run before spans leave the user's process. ML-based detection is planned, so teams needing semantic or adaptive redaction should not treat the current option as a complete privacy system. Granular role-based access control is available. Production deployments can be configured for EU-region processing, with EU-only storage confirmed in the contract during onboarding. Traccia says SOC 2 Type II certification is in progress, and it does not provide a signed Business Associate Agreement; enabling its HIPAA module does not make an organization HIPAA compliant.

Pricing

Hobby costs nothing, with 50K events and 7 days of retention. There is no overage option, making it appropriate for prototypes or low-volume experiments but restrictive once traffic grows.

Observe is $99 a month, including 500K events, 30 days of retention, and $12/100K overage. It suits a small production service that mainly needs tracing and replay. Govern costs $299 a month for 2M events, 90 days of retention, and $8/100K overage, a better fit when runtime controls and longer investigative history matter. Scale is $799 a month for 10M events, one year of retention, and $5/100K overage; high-volume teams get a lower incremental event rate and substantially longer history.

Enterprise uses custom pricing, with volume-based included events, custom overage, and 10 years of retention. The plans are event-based, so teams should estimate agent activity before choosing a tier. Product support for accounts, activation emails, integrations, and the hosted platform is available at [email protected].

Platforms

Traccia is offered through an API and web interface, with hybrid deployment options and an SDK-based agent framework model. The local SDK export path also allows telemetry to remain in an OpenTelemetry-compatible environment without an account.

Who it's for

Choose Traccia if your team needs agent-level traces plus controls that can halt calls, and if evaluation before promotion is part of the release process. It is especially relevant to Python or TypeScript teams using the documented frameworks, and to organizations that want EU-region deployment or an existing telemetry backend. It is a poor fit for HIPAA workloads requiring a signed BAA, and less compelling if you only need basic tracing without governance or evaluation.

Pros and cons

• Pros: Runtime spend, model, and loop controls can prevent calls; traces include sessions and tool activity; evaluations support datasets and experiments; local OpenTelemetry export avoids account dependence.
• Cons: Hobby retention is only 7 days; paid plans impose event quotas and overage charges; automatic cost tracking is specifically described for OpenAI and Anthropic; ML PII detection is not yet available; SOC 2 Type II remains in progress; no signed BAA is offered.

Alternatives

Pick SigNoz instead if self-hosting across Android, API, iOS, web, and other environments is more important than Traccia's agent policy controls. Agenta is a stronger alternative for teams wanting a free hobby allowance of 5,000 agent runs per month and 20 evaluations, with one-week trace retention. Choose PetalTrace when a free, self-hosted tool spanning Linux, macOS, Windows, API, and web is the priority.

TraceRoot starts at $30 a month with unlimited seats, 150k events, and 30-day retention, making it attractive for seat-heavy teams. VoltOps offers a free developer plan with one seat, one prompt, one project, and 250 traces per month. Arize AX provides a free SaaS tier with 25k trace spans per month, 1 GB ingestion, 15-day retention, unlimited users, and unlimited evaluations.

Braintrust is worth considering for its free starter allowance of 1 GB processed data, 10,000 scores, unlimited users and projects, and 14-day retention. Fiddler AI is an alternative for teams focused on real-time guardrails and AI observability. See the broader AI Agent Observability Tools list for additional options.

Verdict

Traccia is a strong choice for teams that need agent tracing tied directly to runtime limits and pre-release evaluation. Its main reason to choose it is the combination of intervention and observability, including an account-free local export path. Look elsewhere for HIPAA programs requiring a signed BAA, for mature certification today, or for simpler tracing with fewer event and retention trade-offs.