LanceDB, Pinecone, Qdrant, Chroma, turbopuffer, Weaviate and Zilliz.
With a CLI: Pinecone, Qdrant, Chroma and Zilliz.
Of the 7 vector databases we drove up to 2 Nov 2026, LanceDB, Qdrant, Zilliz and Weaviate work best for AI agents. With them, our agent got in, did nearly all the core work and could leave without a person. Sign-up stopped our agent at 2 of 7.
7 vector databases, driven 26 Oct to 2 Nov 2026.
Best for vectors stored in the agent's own process
The open-source library installed from npm and PyPI and ran every core task with no account and no key: tables, upserts, deletes, vector, full-text and filtered search, indexes and branches.
Best for a hosted vector database an agent runs by API
Sign-up took an email code, and our agent made one management key in the web console.
Best for an agent that wants a CLI and a REST API for everything
After a sign-up in the browser, clusters, collections, inserts, full-text and hybrid search, imports and access control all worked over REST, the CLI and the SDK.
Best for answers over the data, not only search
After a sign-up with an email code, our agent made a free cluster and a key in the console, then did every core task without a browser: objects, vector, keyword, hybrid and filtered search, Query Agent answers with sources and Engram memories.
Based on 7 of the 246 products we drove, last driven 2 Nov 2026.
| # | Vector database | Level | Find | Get in | Use | Pay | Leave | Core work | In after | Driven | |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | LanceDBRuns locally | L4 Runs on its own | On its own | On its own | On its own | Stopped | On its own | 8 of 8 | 2 min | 30 Oct | 8 of 8 core tasksIn after 2 minDriven 30 Oct |
| 2 | ChromaNo way in on its own | L1 Needs a person | On its own | Stopped | Not reached | Not reached | Not reached | Unknown | Unknown | 31 Oct | Unknown core tasksIn after unknownDriven 31 Oct |
| 3 | PineconeNo way in on its own | L1 Needs a person | On its own | Partly | Not reached | Not reached | Not reached | Unknown | Unknown | 26 Oct | Unknown core tasksIn after unknownDriven 26 Oct |
| 4 | turbopufferNo way in on its own | L1 Needs a person | On its own | Stopped | Not reached | Partly | Partly | Unknown | Unknown | 31 Oct | Unknown core tasksIn after unknownDriven 31 Oct |
| 5 | QdrantSign-up in a browser | L4 Runs on its own | On its own | On its own | On its own | Partly | On its own | 7 of 7 | 4 min | 2 Nov | 7 of 7 core tasksIn after 4 minDriven 2 Nov |
| 6 | ZillizSign-up in a browser | L4 Runs on its own | On its own | On its own | On its own | Partly | On its own | 7 of 7 | 6 min | 28 Oct | 7 of 7 core tasksIn after 6 minDriven 28 Oct |
| 7 | WeaviateSign-up in a browser | L4 Runs on its own | On its own | On its own | On its own | Partly | On its own | 10 of 10 | 8 min | 28 Oct | 10 of 10 core tasksIn after 8 minDriven 28 Oct |
7 vector databases, driven 26 Oct to 2 Nov 2026.
AgentDrives StudyWhich vector database works best for AI agents?
Products in Vector databases at L4 or higher.
57%4 of 7
| What | Count | Out of |
|---|---|---|
| Products in Vector databases at L4 or higher. | 4 | 7 |
AgentDrives drives, 1 Oct – 18 Nov 2026
Cite this figure
AgentDrives, "Products in Vector databases at L4 or higher.", 18 Nov 2026. https://agentdrives.com/research/best-vector-databases-for-ai-agents#dc-1
Didn't get in on its own: Chroma, Pinecone and turbopuffer.
7 vector databases, driven 26 Oct to 2 Nov 2026.
LanceDB, Pinecone, Qdrant, Chroma, turbopuffer, Weaviate and Zilliz.
With a CLI: Pinecone, Qdrant, Chroma and Zilliz.
LanceDB, Qdrant, Weaviate and Zilliz.
First key made in a browser: Qdrant, Weaviate and Zilliz.
LanceDB, Qdrant, Weaviate and Zilliz.
Runs on its own (L4 or higher): LanceDB, Qdrant, Weaviate and Zilliz.
None of the 7.
First stop: payment: LanceDB, Qdrant, Weaviate and Zilliz.
LanceDB, Qdrant, Weaviate and Zilliz.
Not reached: Chroma and Pinecone.
The open-source library installed from npm and PyPI and ran every core task with no account and no key: tables, upserts, deletes, vector, full-text and filtered search, indexes and branches. The hosted side is closed to agents: LanceDB Cloud takes no new sign-ups, and its REST API comes with Enterprise, sold through a sales form.
Chroma is a database that stores data for AI search. The agent read all the public docs and specs, but could not make an account. The sign-up answered 'sign-ups are temporarily disabled' every time, and Chroma says its cloud service is winding down.
Pinecone is a vector database that stores data so software can search it by meaning. Our agent could read the documentation and list the hosted models on its own, but it could not create an account. The email sign-up showed a Cloudflare 'Verify you are human' checkbox.
turbopuffer is a database for searching text and vectors. An agent could create an account with an emailed code, but could not get an API key because the key sits behind a paid card subscription with no free plan. The agent got as far as the card form and did no database work.
Sign-up took an email code, and our agent made one management key in the web console. Everything after that ran without a browser: it created a free cluster over the Cloud API and did every core task, server-side embeddings included. On the hosted cluster, filtered search and delete-by-filter first need a payload index.
After a sign-up in the browser, clusters, collections, inserts, full-text and hybrid search, imports and access control all worked over REST, the CLI and the SDK. The documented API key endpoints answered with errors, so the first key comes from the console or a device login confirmed in a browser.
After a sign-up with an email code, our agent made a free cluster and a key in the console, then did every core task without a browser: objects, vector, keyword, hybrid and filtered search, Query Agent answers with sources and Engram memories. New free clusters refuse the deprecated GraphQL API, which its docs still show in places.
Our agent drove each vector database between 26 Oct and 2 Nov 2026: find the docs, get in, do the core work, pay, leave. It signs every request, never solves a CAPTCHA and signs up with its own address. Our method
Level first, then the share of core work done, then minutes to get in. Levels run from L0 to L5. What the levels mean
AgentDrives Lab. “Which vector database works best for AI agents?” 3 Nov 2026. agentdrives.com/research/best-vector-databases-for-ai-agentsDrive data CC BY 4.0
No. A Cloudflare Turnstile check kept our agent out at sign-up, so it never had an API key. Without one it could only search Pinecone's docs and list its hosted models.
All 7 publish one. Our agent did real work through the servers of Qdrant, Zilliz and Weaviate. LanceDB's only searches its docs, and the servers of Chroma, Pinecone and turbopuffer need an API key it couldn't get.
At the checkouts our agent reached: turbopuffer $16 a month, Qdrant $26.10 a month, Weaviate $45 a month and Zilliz $99 a month, each a card form. Qdrant, Weaviate and Zilliz have a free plan before that; turbopuffer has none. LanceDB sells its hosted plan through sales only.
Not yet: 0 of 7. Each takes a card on a checkout form, and none offers an agent payment such as x402.