Create workspace
Accept your pilot invitation and create a workspace in the hosted console.
Agent visibility · Control · Privacy
Trace agents. Guard tools.
Protect personal data.
TracerYX brings agent tracing, explicit tool guards and signed evidence together with Cloak, our local PII redaction and tokenization product.
Read the Cloak benchmark report ↗Agent platform
Trace. Guard. Review.
Python and TypeScript SDKs, explicit tool policies and a hosted console for traces and signed evidence.
Explore agent workflows ↗TracerYX Cloak
Redact. Tokenize. Locally.
Detect personal identifiers and replace them with opaque tokens in supported traces before export.
Explore local privacy protection ↗The TracerYX agent platform
Trace your AI agents. Guard tools. Review evidence.
TracerYX helps teams trace instrumented agent activity, apply policies before guarded tool calls execute, and review signed evidence.
Python and TypeScript SDKs run alongside your agent code; the hosted console provides the dashboard and evidence records.
Pilot signup currently requires an invitation. Contact info@traceryx.com about access.
Python + TypeScript SDKsExplicit tool guardsSigned evidence
Illustrative evidence record
ExampleInspect recorded activity.
03
send_email(to=████@███.com)redacted07
record_event(value=tok_████)redacted04
policy: guarded_tool_checkpassData lineage
Inspect activity you instrument.
Trace instrumented activity and review policy outcomes for tools you explicitly guard with the SDK. TracerYX does not automatically intercept every tool call or network request.
lookup_orderrecorded
send_emailallowed
export_recordsblocked
3 example toolsGuarded calls: allowed or blockedCoverage depends on integration
How it works
From workspace to reviewed evidence.
Start with an invitation, then connect your agent code to the hosted console. Follow the public docs for SDK setup, instrumentation, tool guards, and evidence verification.
Create API key
Create an API key for your workspace to connect your agent instrumentation.
Install SDK
Install the Python or TypeScript SDK alongside your agent code.
Instrument and guard tools
Add instrumentation and explicitly integrate SDK guards before outgoing tool calls execute.
Run agent
Run your instrumented agent with your configured policies.
Inspect traces and verify evidence
Use the hosted console to inspect recorded activity and review signed evidence.
Clear evidence, fewer review loops.
The hosted console brings recorded traces, guarded tool policy outcomes, and signed evidence records into your review workflow.
The agent SDK’s supported telemetry redaction happens in the agent process before transmission. Its detection is pattern-based and does not guarantee detection of all PII, PHI, or secrets. Telemetry redaction is separate from guarding outgoing tool calls, which requires explicit SDK integration.
Trace
Instrumented agent activity. review what your integration records.
Policy
Policy outcomes for guarded tool calls. inspect the checks applied before execution.
Evidence
Signed evidence records. verify integrity using the documented verification process.
Evidence and its limits
Verify integrity. Review the context.
Signed evidence helps you check the integrity of a recorded artifact. Signatures do not prove that recorded activity is true or complete.
Signed evidence
A record you can inspect
Review the recorded activity and policy outcomes, then follow the verification documentation to check evidence integrity.
Integration matters
Coverage starts in your code
Instrumentation determines what is recorded. Explicit SDK guards apply policies before guarded tools execute. Hardware attestation and confidential compute are not currently shipping capabilities.
TracerYX Cloak / Local processing preview
Protect personal data.
Keep useful context.
Cloak is our dedicated PII redaction and tokenization product. Its first-party detector processes supported traces locally before they leave your environment. Evaluate it for your own data path alongside TracerYX’s agent platform.
The identifiers change. The story remains.
Synthetic trace example · Token IDs abbreviated
Before Cloak Inside your environment
customer_name: "Jordan Patel"
user_email: "jordan.patel@example.com"
backup_email: "jordan.patel@example.com"
customer_id: "CUST-99821"
handling_note: "Keep it refrigerated"After Cloak Ready for your collector
customer_name: "[PERSON:…]"
user_email: "[EMAIL:same-token]"
backup_email: "[EMAIL:same-token]"
customer_id: "[ACCOUNT:…]"
handling_note: "Keep it refrigerated"Illustrates behavior verified in our installed SDK example. Aliases are ephemeral pseudonyms, not a reversible vault; new batches receive new aliases. Detection can still miss identifiers or remove useful text.
Privacy starts where the data lives
Your environment.
Your privacy controls.
Put redaction between your application and your telemetry destination. Keep raw trace text within your environment while sending a protected representation onward.
Detect in context
First-party detection combines trained models, contextual corrections and deterministic checks to identify personal information in supported trace attributes.
Replace with tokens
Replace detected identifiers with opaque aliases. Repeated exact values of the same type share an alias within a call or batch, so related activity stays connected.
Protect before export
Process supported traces inside your environment. The local detector runs without a remote inference API, before the protected payload reaches your collector.
Your applicationCloak · local processingYour collector
Measured. Documented. Open to scrutiny.
The evidence behind Cloak.
Our latest retained detector was evaluated alongside Presidio, OpenMed, NVIDIA GLiNER, Gravitee BERT and Open Privacy Filter. Explore the comparisons, datasets and tradeoffs in the release report.
95.79%
Complete-span coverage
90.36%
Character F1
790
Annotation-valid documents
Latest English Nemotron + Gretel cohort: 797 retained documents, 790 with valid annotations. Complete-span coverage requires every annotated character in an identifier to be protected. Character F1 balances coverage and precision. Results depend on the dataset and operating point.
28.19 percentage points higher complete-span coverage than the tested Presidio setting.
Cloak: 95.79% at threshold 0.20. Presidio: 67.60% at 0.01. OpenMed achieved higher character F1 and lower outside-annotation removal in this cohort. These comparisons do not establish universal superiority.
Built for a real integration
Choose the balance.
Understand the boundaries.
Adjust detection sensitivity
Choose High coverage, Balanced or Preserve text. A lower threshold can catch more identifiers and also remove more ordinary text. Mandatory deterministic findings remain active.
See the measured tradeoffs ↗Preserve trace structure
The qualified Node.js 22 integration handles OTLP HTTP/JSON traces. Validated structural trace and span IDs and timestamps are preserved while supported attributes are processed.
Fail without raw fallback
Malformed or unsupported inputs are rejected. Processing and export failures do not trigger a fallback that forwards raw trace data. Values classified as secrets or payment data are discarded.
Evaluate on your data
Benchmark performance is a starting point. Check detection and text preservation on representative traces before rollout. The current integration does not support logs, metrics, protobuf or gRPC.
Start with your workflow
Bring your agents into view.
Put privacy in the data path.
Explore the TracerYX agent platform or evaluate Cloak locally. Contact our team about pilot access, integration and the right controls for your workload.
Cloak is available as a local processing preview. Commercial distribution requires completed release approvals and operator configuration. See the release report for availability and scope.