Saiku 4.8 puts one cube behind a browser, a typed REST surface, and an MCP endpoint, and its own Dockerfile disagrees with the README about where MCP lives
Open-source semantic layer: one cube for Excel (MDX/XMLA), dashboards, and AI agents (MCP). Mondrian + Apache Calcite.
At a glance
- What is it?
- Spicule rebuilt the Mondrian OLAP browser as a semantic layer with two SQL planners, two parallel AI API families sharing one validation envelope, admin editable skills and spaces, and a demo image that refuses to boot on the default password once it is reachable. The interesting parts are the guards, and the two places where the documentation and the build file contradict each other.
- Who is it for?
- Saiku is worth a look if you have Mondrian cubes already and want one of them to answer both a drag and drop user and an agent through the same contract.
- Can I use it commercially?
- Yes. Apache-2.0 is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository received new commits within the last day.
- What is it written in?
- Mainly Java, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.
Star history
- spiculedata/saiku1,339 ★ · +26
| Repository | 2026-08-02 | 2026-11-15 |
|---|---|---|
| spiculedata/saiku | 1.3k | 1.3k |
26 Stars in the last 105 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
The demo boots on admin/admin and refuses to boot on it once reachable
The whole evaluation path is one command:
docker run -d -p 8080:8080 --name saiku -e SAIKU_DEMO=true ghcr.io/spiculedata/saikuOpen http://localhost:8080/ui/ and the login is `admin` / `admin`. Demo mode ships a self contained H2 plus FoodMart cube, so dragging fields onto rows, columns, or filters makes the SPA write the MDX for you. That default is intentional and fenced. For a real deployment you either drop `SAIKU_DEMO=true` or set an admin password with `-e SAIKU_ADMIN_PASSWORD='a-strong-password'`, because Saiku refuses to start on the default pair once it is network reachable. A demo credential that is also a production guard rail is a good pattern, and the one thing to notice is that the port mapping is what makes the container reachable, so the same command with the mapping removed is the difference between a working demo and a process that will not start.
A fresh volume works, an old one needs a chown, and the failure mode refuses to rotate keys
The container runs as a non root user at uid and gid `10001:10001`. A fresh named or anonymous volume is fine out of the box. A pre existing `saiku-home` carried over from an older root container, whether a bind mount or a named volume, has to be re owned once with `sudo chown -R 10001:10001 <host-dir>`, or for a named volume through `docker run --rm -v saiku-home:/app/saiku-home --user 0 --entrypoint chown ghcr.io/spiculedata/saiku:latest -R 10001:10001 /app/saiku-home`. On Kubernetes the same fix is `securityContext: { runAsUser: 10001, fsGroup: 10001 }`. What makes this a design decision rather than an annoyance is the alternative the authors rejected: the container fails closed with a `FATAL:` message that names the fix instead of silently rotating its encryption key. A rotated key on a home directory holding saved queries and encryption material is quiet data loss, so the loud crash is the better of the two behaviours.
The base image is pinned by digest and the jar arrives pre built
The Dockerfile builds from `eclipse-temurin:21-jre-noble` with a sha256 digest appended, and the comment explains why both are kept: the digest is what actually resolves, so a re tagged or tampered upstream tag cannot change what gets built on, while the tag stays for humans reading the file. Rolling it forward means resolving the new multi arch index digest with `docker buildx imagetools inspect eclipse-temurin:21-jre-noble` and updating both. The OpenTelemetry Java agent is pinned the same way, with an `OTEL_AGENT_VERSION` and an `OTEL_AGENT_SHA256` build argument. Maven never runs at image build time. The jar is produced by the launcher job, staged into `./build-context/saiku.jar`, and copied in, which the comment says avoids needing GitHub Packages auth inside the container build. The runtime user note also fixes the numbers on purpose: existing homes have already been chowned to 10001:10001, so changing the uid without a migration note breaks every deployment.
Calcite plans the SQL by default and the legacy builder sits behind a system property
Saiku began in 2010 as an open source OLAP browser for Mondrian and was rebuilt in 2026 as a semantic layer. The engine underneath is Mondrian 4.8.1.x from a Spicule fork, carrying two SQL planners: a Calcite based one and the legacy `SqlQuery` builder. Calcite is the default, and forcing the old path is a single flag, `-Dmondrian.backend=legacy`. Cellsets go over the wire as Apache Arrow, so the browser and any programmatic consumer share one result envelope rather than two encodings. The reason the Calcite path matters is reach: the documented walkthrough in `examples/lakehouse-demo` goes Saiku, then Mondrian, then Calcite, then Trino, then Iceberg, which is a very different destination set from the OLAP engines the legacy builder was written for. Cells also come back typed as `{value, formatted, unit}`, so a consumer can tell a currency from a count without parsing a string.
Two AI API families share one list, schema, query, and error contract
The agent facing surface is `/rest/saiku/api/ai/*`. `/ai/cubes` and `/ai/schema` make hierarchies, levels, measures, and synonyms discoverable, and a single `POST /ai/query` takes a JSON description of a question, compiles it to validated MDX, runs it, and returns the typed cells. Every validation failure carries a `{status, field, available}` envelope, which is the part an agent actually needs: it learns which field was wrong and what was allowed without reading a log. Cubes also describe themselves through a `saiku.semantic.*` annotation namespace, so synonyms and hints travel with the model rather than in a separate prompt you maintain by hand. There is a second family at `/rest/saiku/api/ai/ossie/*` for datasets modelled with the Ossie semantic layer, with the same list, schema, and query shape, the same `VALIDATION_ERROR` envelope, plus `POST /ai/ossie/ask` for a plain language question. Anomaly and forecast endpoints at `/ai/anomaly` and `/ai/forecast` share the envelope too. The structural catch is that these are two semantic layers with two endpoint families, so a cube modelled the old way gets the MDX side and the Ossie endpoints have nothing to work on.
The README ships a stdio wrapper that the Dockerfile says no longer exists
The documentation tells you the container bundles `saiku-mcp`, a stdio Model Context Protocol wrapper, and shows the entire client configuration:
{
"command": "docker",
"args": ["exec", "-i", "saiku", "saiku-mcp"]
}The Dockerfile header says something else. Its comment states that since issue 878 the MCP endpoint lives inside saiku-webapp at `/rest/saiku/api/mcp`, that MCP hosts such as Claude Desktop and Cursor authenticate with per user Basic credentials over the same chain as the AI REST API, and that no separate stdio binary is required. So the one line of config a user is told to paste calls a binary the build file says is gone, and it carries no credentials for an endpoint the build file says sits behind an auth chain. Either the wrapper bridges the auth internally or the snippet is out of date. Either way, the first thing to verify before wiring an agent to this is which of the two shapes 4.8 actually ships.
Skills and personas are files an admin drops into the same volume they chowned
The AI surface can be extended without writing code. Skills are markdown workflows with YAML frontmatter under `saiku-home/skills/*.md`, discoverable from `/ai/ask`; you invoke one by prefixing an ask with `/<skill-name>`, or you ask in natural language and let the model route through the skill catalogue. Spaces are JSON personas under `saiku-home/agent-spaces/*.json` that bind an ask to a system prompt, a cube allowlist, and a skill allowlist. The enforcement detail is the part worth knowing: `POST /ai/spaces/{id}/ask` applies the persona server side, so a cube outside the allowlist comes back as a 403 the user cannot override from the client. That is the right place for the boundary. It also means the AI behaviour of an instance is editable by anyone who can write into the volume that the previous section told you to chown by hand.
Tracing is opt in through one env var and quality floors are files gated on every pull request
Observability is attached without touching code. Setting `OTEL_EXPORTER_OTLP_ENDPOINT` loads the OpenTelemetry Java agent and auto instruments Jetty, Jersey, JDBC so every piece of SQL Mondrian emits becomes a child span, outbound HTTP, JVM metrics, and the DBCP2 pool. Without that variable the agent is never loaded at all, so the overhead claim is structural rather than measured. Trace context is injected into the Saiku log pattern. The gap is named too: Tier 2 custom spans for services such as `ThinQueryService` are listed as not yet covered. Quality is handled the same declarative way, with per module test count and line coverage floors plus UI type checks declared in `.github/test-floors.json` and `.coverage-thresholds.json` and gated on every pull request by the ci workflow, alongside a separate weekly run whose workflow file the description stops naming halfway through `quality-report.ym`.
Editorial conclusion
Saiku is worth a look if you have Mondrian cubes already and want one of them to answer both a drag and drop user and an agent through the same contract. Check three things before you commit: whether your cubes live on the Ossie semantic layer or on plain Mondrian, since the two AI API families are separate, and if it is the latter then the Ossie endpoints do nothing for you; whether you can hand over Basic credentials to your agent runtime, because the MCP path inherits the AI API auth chain; and whether you are prepared to chown a data volume to 10001:10001 and keep an eye on a base image pinned by digest. For a hosted evaluation, the demo at demo.saiku.bi resets nightly, which is the cheapest way to see whether the cube contract survives contact with your model.
Frequently asked questions
How do I run Saiku locally and what are the demo credentials?
Run the image with SAIKU_DEMO=true, open http://localhost:8080/ui/, and log in with admin / admin. Demo mode ships a self contained H2 and FoodMart cube so you can drag fields onto rows, columns, or filters right away.
What happens if I leave the default Saiku admin password in place?
Saiku refuses to start on the default admin/admin once it is network reachable. Either drop SAIKU_DEMO=true for a real deployment or pass -e SAIKU_ADMIN_PASSWORD='a-strong-password'.
Do I need to chown an existing Saiku data volume?
Only if it already exists. A fresh named or anonymous volume works out of the box, while a saiku-home carried over from an older root container must be re owned to 10001:10001 once. The container then fails closed with a FATAL message naming the fix rather than silently rotating its encryption key.
How do I switch Saiku back to the legacy SQL builder?
Start it with -Dmondrian.backend=legacy. The Calcite based SQL planner is the default and sits alongside the legacy SqlQuery builder in the Mondrian 4.8.1.x Spicule fork.
How do AI agents query Saiku?
Through the typed REST surface under /rest/saiku/api/ai, where /ai/cubes and /ai/schema expose hierarchies, levels, measures, and synonyms, and POST /ai/query turns a JSON question into validated MDX and returns cells typed as {value, formatted, unit}.
Can I extend Saiku's agent behaviour without writing code?
Yes. Skills are markdown files with YAML frontmatter under saiku-home/skills, and Spaces are JSON personas under saiku-home/agent-spaces that scope an ask to a system prompt plus cube and skill allowlists, enforced server side so a disallowed cube returns a 403.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/spiculedata-saiku)Star history chart for your README
[](https://hysenlabs.com/projects/spiculedata-saiku)Was this useful?