Attu is the open-source administration GUI for Milvus. It gives you a browser console for designing schemas, managing collections and databases, browsing data, running vector searches, managing users and roles, and watching cluster status. If you run Milvus 2.6.0 or later, or use Zilliz Cloud, run Attu 3.x. The 2.6.x line is no longer maintained, and version 2.6.5 has unpatched Playground flaws that let an unauthenticated attacker proxy requests to internal services. Attu is worth deploying for exploring data, administering several clusters and handling permissions by eye. Keep it on a private network, and don’t treat it as a replacement for the SDK in application code or for a real alerting stack.
Is Attu part of Milvus itself, and what license does it ship under?
No. Attu is a separate tool, not a component built into the database. The Milvus documentation calls it “an open-source GUI tool for intuitive Milvus management”. The tool that is built into Milvus is the Milvus Web UI. The Milvus docs describe it as a built-in observability page, and they set it against Attu and Birdwatcher to show the difference. In practice, Attu ships as its own container image (zilliz/attu) with its own Milvus version compatibility rules.
Both the Milvus docs and Alibaba Cloud’s managed Milvus docs describe Attu as open source. The Milvus Operator and Milvus CLI repositories both show Apache-2.0, but that tells you nothing about Attu. Before you bundle Attu into a product or an internal platform, read the licence file in the Attu repository itself. Licensing for the database is a separate question, covered in our explainer on Milvus’s licence, governance and cloud deployment options.
Does Attu work with managed Zilliz Cloud, or only with self-hosted Milvus?
Both. The official quickstart says Attu connects to Milvus Standalone and to Zilliz Cloud, “providing flexibility to work with local or cloud-hosted databases”. The Attu 3.0 release notes list Zilliz Cloud as a supported target alongside Milvus 2.6.0+ and Milvus 3.x. Some managed Milvus offerings also integrate Attu themselves. Alibaba Cloud’s Milvus service, for example, offers Attu on port 3000 at both an internal and a public endpoint per instance.
What can you actually do in Attu, and what does it not cover?
Attu covers Milvus administration from schema design to monitoring. Zilliz’s feature list includes:
- visual design, viewing and editing of collection schemas
- management of multiple databases and collections, including cloning
- a data browser with filters, inline editing and import/export
- interactive vector similarity search with visualisation, expression-based queries and a built-in RESTful API editor
- visual management of RBAC and privilege groups, real-time monitoring of nodes, segments and tasks, and slow-query analysis
Version 3.0 is a complete rewrite. It adds multi-cluster management, persistent local state, GUI-based backup and restore, an API Playground and built-in Prometheus metrics. The metrics dashboard shows 16 or more core metrics, and you set a Prometheus URL for each connection. Attu 3.0 also handles External Collections over Parquet and Iceberg data, snapshots, TEXT fields with BM25 workflows, and online schema evolution. It also has aggregation queries (COUNT, SUM, AVG, MIN, MAX), Function Score reranking and MinHash-based deduplication.
The 3.0 release also adds a built-in AI agent connected to more than 50 Milvus tools. Attu’s documentation stresses that the agent is not a documentation chatbot: it inspects cluster state and carries out real operations through Attu. Attu doesn’t bundle an LLM. You bring your own provider (OpenAI, Anthropic, DeepSeek, Google Gemini, OpenRouter or any OpenAI-compatible endpoint), and the API key is encrypted locally rather than uploaded to a service Attu runs.
Some things are out of scope. Attu is an interactive console, so it isn’t where your ingestion pipelines, schema migrations or application queries should live. Those belong in code, versioned and reviewed, using an SDK. Its Prometheus dashboard shows metrics, but nothing we found describes alert routing or on-call notification. Alerting stays with your Prometheus and Alertmanager (or equivalent) setup, and Attu is a second place to look, not the system that wakes someone up.
Running Attu next to Milvus with Docker, Docker Compose or Helm
Attu 3.0 keeps a local database of cluster configs, agent history, custom skills, LLM settings and user preferences. Without a mounted volume, a container restart wipes the lot. Zilliz’s own example is:
docker run -d --name attu -p 3000:3000 -v attu-data:/data zilliz/attu:v3.0.0-beta.6
That command is from the beta announcement. Version 3.0.0 was released on 3 September 2026, so pin a checked 3.x release tag rather than a beta.
There are also options without containers. Attu 3.0 comes as desktop apps for macOS, Linux and Windows, and as a standalone Linux x64 server tarball for running it without Docker or Electron. The Docker and standalone builds bundle milvus-backup for backup and restore support.
Adding an Attu service to a Milvus Docker Compose file
Translating the docker run command above into Compose is straightforward. Add this next to your existing Milvus standalone services:
services:
# existing etcd, minio and standalone Milvus services stay as they are
attu:
image: zilliz/attu:v3.0.0 # pin the 3.x tag you have verified
ports:
- "127.0.0.1:3000:3000" # loopback only; reach it via SSH tunnel or VPN
volumes:
- attu-data:/data
restart: unless-stopped
volumes:
attu-data:
Merge attu-data into the top-level volumes: block if your file already has one.
Enabling Attu in the Milvus Helm chart
The Milvus Helm chart includes an attu block, which is switched off by default (enabled: false). The chart’s default image tag is v2.5.3. That predates the 3.0 line Zilliz now points users to, so override it when you turn Attu on. The chart’s Attu service defaults to ClusterIP on port 3000, so it isn’t reachable from outside the cluster until you add an ingress:
attu:
enabled: true
image:
tag: v3.0.0
service:
type: ClusterIP
port: 3000
To reach it, run kubectl port-forward to that service, or use an internal-only ingress.
What about the Milvus Operator?
The Milvus Operator README recommends the Operator for production, citing managed rolling upgrades, dynamic configuration reload and health diagnostics. It recommends Helm for beginners. The README describes how the Operator manages the Milvus stack. It doesn’t describe an Attu component. With the Operator, we suggest running Attu as its own small Deployment and Service using the same zilliz/attu image, with a persistent volume at /data, and connecting it to the Milvus service inside the cluster.
Logging in to Attu when Milvus authentication is on
Attu logs in with a Milvus username and password. So the first step is turning authentication on in Milvus itself by setting common.security.authorizationEnabled: true. How you set it depends on how you deploy:
| Deployment | Where authorizationEnabled: true goes |
|---|---|
| Docker Compose | the Milvus config file (common.security) |
| Helm | extraConfigFiles → user.yaml in values.yaml |
| Milvus Operator | spec.config.common.security in the Milvus custom resource |
Milvus creates a root user with the password Milvus when the instance is first initialised. That user holds the built-in admin role, which can access everything. In Attu’s connection dialog, enter the database, the username and the password. SDK clients pass the same pair as a token in the form root:Milvus. Without a valid token, the connection fails with a gRPC error. Change the root password before anything else. Milvus supports this through the SDK’s update_password call, and a configured super user can reset a forgotten password.
Attu 3.0 adds a login layer of its own. The app runtime now has local authentication with user isolation and an 8-character minimum password, plus an admin view of all saved connections and an audit log.
Which Attu version should you run for your Milvus version?
Start from what Zilliz still maintains. Attu 3.0 supports Zilliz Cloud, Milvus 2.6.0+ and Milvus 3.x. Zilliz has said the 2.6.x Attu line is no longer maintained and won’t get fixes.
| Your Milvus | Attu to run | Notes |
|---|---|---|
| Milvus 3.x | Attu 3.x | External Collections and TEXT fields also need Storage V3. |
| Milvus 2.6.x | Attu 3.x | Supported. External Collections, snapshots, TEXT fields, MinHash, entity-level TTL and some schema operations need Milvus 3.x. |
| Zilliz Cloud | Attu 3.x | Listed as supported. |
| Milvus older than 2.6.0 | No maintained Attu line | Attu 3.0 needs 2.6.0+, and the 2.6.x Attu line is unmaintained. Upgrading Milvus is the clean route. |
Defaults deserve a check. The Helm chart’s v2.5.3 tag is one example of a pinned version drifting behind what the vendor supports, and an enabled: true added on top of it quietly deploys an old console. Attu 3.0 also manages schema actions based on the connected Milvus version, and some schema operations require Milvus 3.x.
Can one Attu instance manage several Milvus instances or databases?
Yes, from 3.0 onwards. Attu 3.0 lists every Milvus connection in one sidebar, with a health dot beside each (green for reachable, red for down), and switching between clusters takes one click. In 2.x you had to disconnect and reconnect. Inside each cluster, Attu manages multiple databases and their collections. Each connection gets its own agent session and its own Prometheus URL. Administrators can see and export the saved connections from an admin view.
That convenience means one console can now hold connections to development, staging and production. Before you connect production to a shared Attu instance, decide who is allowed to use that instance.
Managing Milvus users, roles and privileges through Attu
Attu is a front end to Milvus’s own role-based access control (RBAC). Understanding that model is what makes the screens make sense. Milvus RBAC has four parts. A resource sits at one of three levels: instance, database or collection. A privilege is an operation, such as creating collections or inserting data. A privilege group bundles several privileges. A role combines privileges with the resources they apply to. Users get roles, and one role can go to many users. Changing a role’s privileges changes them for every user who holds it.
The order of work follows Milvus’s RBAC procedure:
- Log in as
root(or another admin) and create the users who need access. - Create roles that match real jobs, such as a read-only analyst or an application writer for one database.
- Optionally, define privilege groups so a role can be granted a bundle in one step.
- Grant privileges or privilege groups to each role, scoped to the instance, a database or a collection.
- Grant the roles to users.
Alibaba Cloud’s Attu docs describe creating users and roles and granting roles to users, and Zilliz lists visual management of privilege groups. You can also manage a user’s permissions per database and collection. Collection actions such as schema design and data import need collection management permissions.
What are the access-control and exposure risks of running Attu in production?
The documented risk is unauthenticated network access to Attu 2.6.5. Bishop Fox found two flaws in Attu 2.6.5. The Playground feature lacked authentication, so HTTP and HTTPS requests to public URLs could be proxied through it. Weak input checks let those requests reach private IP addresses as well. In an EKS cluster running Milvus and Attu, Bishop Fox chained the two to gain administrative access to Attu’s Kubernetes namespace. The advisory warns that an unauthenticated attacker with network access could reach internal services and cloud metadata endpoints, and possibly the wider cloud environment. Version 2.6.5 will not be patched. Bishop Fox couldn’t reproduce either flaw in 3.0.0 and recommends upgrading as soon as possible.
At least one published deployment guide does the opposite. A Milvus blog walkthrough for EKS enables Attu behind an AWS load balancer set to internet-facing “so that Attu can be accessed via the Internet”. Managed offerings can also give you a public Attu endpoint. Our position is simple. Treat Attu like a database admin port. Keep it on ClusterIP or loopback, reach it through a VPN, SSH tunnel, kubectl port-forward or an internal-only ingress, and never put it on a public address, whatever version you run.
Other risks come from the access Attu holds. Attu persists the cluster configs for every connection it manages. Its AI agent carries out real operations. Its Playground proxies HTTP requests, as the 2.6.5 flaw showed. Give Attu connections the least-privileged Milvus role that fits the job, keep the default root password out of production, and review Attu 3.0’s audit log.
When should a team use Attu versus the SDK, CLI, or a proper metrics and alerting stack?
Each tool has its own job, and they overlap less than it seems.
| Need | Right tool | Why |
|---|---|---|
| Application reads, writes, schema migrations | SDK (pymilvus, Java, Go, Node.js, REST) | Repeatable, reviewable, testable code |
| Terminal admin, import/export | Milvus CLI | Built on the Python SDK; covers connections, data operations and import/export; installs with pip or as a Docker image |
| Quick check of component health and config | Milvus Web UI | Built into Milvus since v2.5.0 at :9091/webui, for system observability, collection details and configuration |
| Exploring data, checking search results, designing schemas | Attu | Visual search, data browser, schema editor |
| Many clusters, permission debugging, GUI backup and restore | Attu 3.x | Multi-cluster sidebar, RBAC workflows in context, backup in the GUI |
| Alerts and on-call | Prometheus with Alertmanager or equivalent | Attu shows metrics; we found no account of it routing alerts |
Zilliz describes the point of GUI backup and restore as taking a workflow that used to need the CLI and making it a visible recovery path for local testing and staging checks. That also marks where Attu stops. A restore you plan to rely on in production should be scripted and rehearsed, and the GUI is for running it by hand or checking it.
Our rule of thumb follows from the table. A single developer on one local Milvus can do without Attu, though it makes inspecting vectors and search results much quicker. A team running several environments, or debugging permissions often, gets real value from Attu 3.x. Either way, the SDK stays the system of record for anything that has to happen the same way twice.