open source · Apache-2.0
Memoturn is an open-source AI engineering platform: tracing, cost and latency metrics, offline, online, and human evals, versioned prompts with deploy channels, monitors, and datasets. OpenTelemetry-native, self-hostable, and nothing leaves your network.
The observability stack the LLM SaaS vendors run — traces, eval pipelines, prompt registries — rebuilt on open infrastructure: Postgres, Apache Doris, Redis, S3. One docker compose away, and your telemetry never leaves your network.
No proprietary cloud, no black boxes. Observe what your app does, evaluate whether it's good, and ship the next prompt: one platform you can inspect, extend, and self-host.
Traces, spans, generations, and scores on a waterfall timeline, with sessions and per-user rollups. Ingest from the SDKs, the OpenAI and LangChain wrappers, or any OpenTelemetry exporter speaking OTLP/JSON with GenAI semantic conventions.
Offline experiments over datasets, online evaluators sampling production traces, and human review queues with one-click scoring. Every score lands in Doris and shows up on the trace it came from.
A versioned prompt registry with production, staging, and custom channels, plus A/B experiments that split live traffic between versions and compare arms by score. Fetch with getPrompt and iterate in a multi-provider streaming playground.
metrics & dashboards
Cost, tokens, and latency (p50/p95) over a Doris rollup, sliced by day, model, user, session, and tool. Pin filtered metrics to custom dashboards and saved views.
monitors & automations
Stateful alert rules and cost budgets evaluated every minute, plus trigger-to-action rules on platform events that fire webhooks or Slack messages when a score drops or spend spikes.
datasets & experiments
Dataset items and experiment runs that link every item to the trace it produced. Benchmark prompts and models side by side in a comparison matrix before anything ships.
embeddings & search
Embedding projections over your traces and semantic find-similar search, so one bad generation leads you to every trace that means the same thing.
playground
Multi-provider with streaming, structured output, and tool calling. Every playground run is recorded as a trace, and any trace payload opens back into the playground.
MCP server
Prompts, datasets, and review queues exposed as tools inside agent IDEs like Claude Code and Cursor, as a local stdio server or over Streamable HTTP per project, with per-tool RBAC.
SSO, RBAC, audit logs, and PII masking are where observability vendors put the paywall. In Memoturn, every one of these ships in the Apache-2.0 core. Self-host it, pass your security review, and pay nobody.
The same traces, scores, datasets, and prompt versions back every workflow, from debugging a single request to benchmarking a model swap, all on infrastructure you own.
Debug production LLM traffic
Follow a request through every span and generation on the waterfall timeline, with full input/output payloads and session grouping across turns.
Track spend by model & feature
Cost, token, and latency rollups per day, model, user, and tool out of Doris. Know what each feature costs before the invoice tells you.
Catch regressions before users do
Online evaluators sample live production traces and score them continuously; monitors turn a bad score or a cost spike into a Slack alert or webhook.
Ship prompts like code
Versioned, immutable prompt registry with deployment channels: promote to production, A/B test challengers on live traffic, roll back instantly.
Benchmark models & prompts
Run experiments over datasets, score them with LLM-as-judge evaluators, and compare runs side by side before switching models.
Self-host for compliance
PII masking at ingest, audit logs, retention policies, and scheduled exports, with every byte of telemetry staying on your infrastructure.
Open source, Apache-2.0. One command locally, docker compose for production, Helm on Kubernetes when you outgrow one host.