Native JSON
Send Ironside’s trace, span, generation, and score model directly when you want explicit control.
Ironside accepts AI interaction data through familiar protocols, preserves the original event evidence, and projects it into a queryable trace system without claiming ownership over evaluation or release decisions.
The raw event remains available even when the query model changes. A projection can be rebuilt; original evidence cannot.
Provider SDKs and evaluation platforms are convenient integration points, but they bring their own schemas and retention assumptions. Once the original interaction is only visible through a derived model, migrations and re-analysis become custom data projects.
Ironside separates intake from interpretation. It writes an immutable raw event log to S3-compatible storage, then processes asynchronous projections into ClickHouse. The trace viewer reads the projection; exports can move the evidence elsewhere. The original event is still there when the query model evolves.
Send Ironside’s trace, span, generation, and score model directly when you want explicit control.
Use OTLP/HTTP and gen_ai semantic attributes with existing instrumentation.
Point a supported client at a familiar API surface while keeping the resulting evidence in your system.
Owner sessions belong to the browser control plane. Instrumented applications use scoped machine credentials with least-privilege access. Projects isolate ownership, while environment values remain observed attributes for filtering—not a substitute for a security boundary.
Original events, normalized projections, trace context, and export state.
A focused trace viewer with project and environment context.
Parquet exports, OTLP forwarding, signed webhooks, and resumable imports.
Evaluation logic, prompt management, or release policy.
Ironside is public source under the Sustainable Use License. Version 0.1.0 is intentionally pre-1.0: the architecture is usable, while interfaces may still change as the system matures.
MinIO, ClickHouse, Postgres, Redis, the API, worker, and viewer are composed as one development stack.
git clone https://github.com/luka-zivkovic/ironside.git
cd ironside
cp .env.example .env
docker compose up --buildExplore the Node SDK ↗