What's new in Grepture

See exactly which team, project, or environment is spending what. Tag requests with a label and Grepture breaks down cost, tokens, and request volume per label.

The OTLP export now ships proper agent traces — span trees, tool-call spans, gen_ai attributes, and cost — plus a one-click example download.

Replay any logged request straight from the traffic log, tweak it, and score the new response against the original side-by-side across your evaluators.

Enforce a tool allowlist at the gateway. Strip disallowed tool definitions before the model sees them, and block or strip disallowed tool calls in the response.

New opt-in debug mode captures the raw input, every redaction, the upstream body, the upstream response, and the restored output for a single request. Auto-deletes after 24 hours.

Per-key and per-label spend rules with email alerts at 50/80/100% and HTTP 402 enforcement at the proxy when the cap is hit.

TypingMind, LibreChat, LobeChat, Open WebUI, NextChat, ChatBox and other BYOK chat clients now route through Grepture with a single base URL change — no plugins. Cost per user, PII redaction, and audit logs across whatever chat app your team already uses.

Traffic logs no longer truncate large request or response bodies. Big payloads — long-context prompts, agent runs with deep tool histories, RAG documents — are stored in full and shown end-to-end in the detail sheet.

Embeddings are persistent surface area, not ephemeral traffic. The new POST /v1/embeddings endpoint redacts PII from the input before OpenAI generates the vector — so PII never lands in your vector store. Own endpoint, own log table, own dashboard tab.

New filter and sort options on the traces view make it easy to spot the most expensive requests and the heaviest token consumers at a glance.

The dashboard now stays snappy on accounts with millions of requests. Analytics, traffic logs, and traces all load dramatically faster.

Tool calls are now a first-class object in Grepture. Track call volume, p50/p95 latency, and error rates per tool across your entire traffic history.

A preview of what we're building next — proxy and dashboard performance improvements, plus ClickHouse-backed traffic logs with dramatically longer retention.

Store LLM provider API keys in Grepture. Automatic same-provider and cross-provider fallback on auth errors, rate limits, timeouts, and server errors.

Keep Pro's low base fee and scale into Business on your own terms. New overage billing smooths the Pro-to-Business gap, and Grepture now bills in USD.

Turn production logs into curated test suites, then run experiments to compare prompt versions and models with LLM-as-a-judge scoring.

Tighter visual identity across the marketing site and documentation with a unified grid layout, improved typography.

Users can now belong to multiple teams, switch between them instantly, and create new teams with their own billing plans.

The trace detail view has been rebuilt with a waterfall timeline, tool call visibility, full-text search, and per-request cost, token, and duration breakdowns.

Tag individual requests with labels, attach key-value metadata, log custom events into traces, and inspect traces in a dedicated detail view.

Split traffic between prompt versions, measure quality with auto-created evaluators, and pick the winner.

Run a local AI gateway in development with the same observability and PII protection as production.

LLM-as-a-judge evaluations that run automatically on production traffic flowing through the gateway.

New trace-only mode gives full dashboard visibility with zero added latency.

Route Cursor AI traffic through Grepture for observability, cost tracking, and PII protection.

Route all Claude Code traffic through Grepture for observability, cost tracking, and PII protection.

Grepture expands from PII redaction proxy to full AI gateway with prompt management, tracing, and more.

Launched Grepture — a programmable proxy that scans every LLM request for PII, secrets, and sensitive data.