Databench Review (2026): Pricing, Features & Honest Verdict

Reviewed by MakerStack · Published · 7 min read

TLDR

Databench is the open-source, self-hosted edition of Alkera, a shared workspace where data teams and an agent work in chats and reactive SQL and Python notebooks. Careful design, but the repo is one day old and the docs trail the marketing. Best for: data teams wanting an agent on their own servers. Price: from $50/mo, free plan yes. Rating: 6.5/10.

What is Databench?

We found Databench on Product Hunt, where it launched on 7 October 2026 as an open-source workspace for data teams. The vendor writes the name as Databench by Alkera. Databench is the Apache-2.0 code you run yourself, and Alkera is the hosted commercial product with the paid plans, which the repo’s own docs call the full offering with many more features.

The idea is simple to describe. You ask an agent a question in a chat, it queries your warehouse, writes SQL and Python into a notebook your teammates can open at the same moment, and it asks before it changes anything. The company behind both is Alkera AI, Inc., a Delaware-law business with a San Francisco mailing address in its privacy policy. This review covers the open-source edition and the hosted plans you would pay for.

What Are Databench’s Key Features?

Shared agent chats in a sandbox

Several people can sit in one chat with the agent at once. In Databench, each chat runs inside a gVisor sandbox on a machine you choose: the bundled local container, or any Linux server you add over SSH. The repo says the default agent harness is a vendored copy of opencode, with a second harness built on Anthropic’s agent SDK listed as work in progress. The marketing site presents both as ready, so treat that second option as unfinished in version 0.5.0.

Reactive SQL and Python notebooks

Notebooks are stored as Python files and follow an existing open-source reactive notebook format, whose engine Databench vendors. A SQL cell can query any database SQLAlchemy supports and hand the result to Python cells as a dataframe. Cells know what they read, so changing one input reruns only the cells that depend on it. The agent edits these notebooks live while you watch, which is the most distinctive part of the product.

A permission gate on every action

This is the strongest design choice. The hosted docs say every tool call is classified by its effect: read, write, destroy, egress, exec or memory. SQL is parsed into a syntax tree, so a DELETE hidden inside a CTE still counts as a delete. You pick a mode from read-only through plan, default and auto to a bypass mode, and you can allow or reject an action once or always. For an agent pointed at production data, that is the right instinct.

Teams, provider keys and your own compute

Databench ships organisations, nested teams, roles, invitations and an authorization layer that records every decision. A model gateway calls Anthropic or OpenAI (Bedrock also appears in the architecture diagram) with your own provider key. The vendor says nothing else leaves your network. Our grep of the repo showed an error tracker wired in but switched off until you set a DSN, and a self-hosted flag that blocks telemetry from phoning home.

What Is It Like to Use Databench?

We did not install it. Going by the setup guide, a trial run is a git clone, a copied settings file with one provider key, and a container build that the vendor warns takes a while on first run. You then sign in on localhost with a seeded admin account. The one-machine stack binds every port to 127.0.0.1 and is meant for your own laptop only.

A team deployment uses a separate production Compose file that refuses to start with development secrets, plus your own TLS proxy and two hostnames. Adding a server is more invasive than most tools admit, and the setup guide is honest about it: it installs packages and gVisor, writes to /opt and /etc, adds an nftables table, edits the Docker forwarding rules and turns on IP forwarding. Servers running ufw or firewalld need manual rules.

Hosted Alkera is lighter: sign up at the web app, verify your email, connect a warehouse and start a chat. Each workspace runs on shared standard compute, and the repo sizes that sandbox at 1 vCPU and 2 GB on Free, 2 vCPU and 4 GB on Plus, and 4 vCPU and 8 GB on Pro. Heavier work moves to org machines charged by the hour. You can also start chats by mentioning the bot in your workplace chat.

Where Does Databench Fall Short?

This repo is brand new. Repository data from the API shows it was created on 6 October 2026 and holds a single squashed initial commit from 7 October, one contributor, no tagged releases and no stars yet. The code itself looks mature, with tests, migrations and a security policy, but there is no public history to judge how bugs get fixed or how often releases ship.

The marketing runs ahead of the docs. Alkera’s site sells dashboards you can share, and the docs page for dashboards says they are coming soon. The lineage docs say a browsable view is also coming soon, so today you ask the agent in chat. The site shows 26 integration logos, including BI and documentation tools, while the docs list 17 data plugins and none of those BI or wiki tools.

Automated pipeline repair, data quality tests and CI gates on the feature pages have no matching docs page that we could find. They also appear to be Alkera features rather than parts of the open-source repo. If those are why you are interested, ask for a demo before you commit.

Security claims need reading closely. The enterprise page says GDPR compliant and lists SOC 2 Type II, ISO 27001 and HIPAA as in progress. The trust center shows engagement letters and monitored controls, not a finished report. Its privacy policy names a card processor and a bot check but no model providers, says nothing about training, and says conversation data can be kept for up to six years unless you delete the account.

One more warning from the docs: workspaces on a dedicated org machine are not isolated from each other, so anyone using that machine, and their agent, can read other workspaces on it.

How Much Does Databench Cost?

Paid plans start at $50 per month for Plus, and Databench itself is free under Apache-2.0 if you host it and pay your own model provider. Hosted Alkera has a Free plan at $0 with 10 GB of storage and limited usage. Plus adds 1 TB of storage, extra usage and access to all models. Pro is $200 per month on the pricing page with 5 TB, much more usage, dedicated support and early access. Enterprise is custom.

Alkera contradicts itself on Pro. Its pricing page says $200 per month, and the billing page in the docs says $250 per month. We would trust the pricing page and confirm at checkout.

There is no annual option. The terms set a one-month subscription period charged automatically each month, and they ask for at least 30 days of notice before the end of a period to stop renewal, while the docs say a cancellation takes effect at the end of the billing cycle and can be undone until then. Nowhere does the pricing page say per seat or flat, but the terms say each licence is for one individual, so budget per person.

Usage runs on credit. Each message draws on a monthly allowance by model, effort and work done, and org machines draw by the hour. Alkera does not publish the dollar size of each plan’s allowance. Top-ups come in $25, $50, $100 or $250 amounts, or any custom amount from $5, and prepaid credit never expires.

For comparison, the best-known hosted notebook rival charges $36 per editor per month for its individual plan and $75 per editor for teams, and another lists $39 per editor per month billed yearly with a free tier of up to 3 editors. Alkera’s $50 sits in between on price, but only Databench gives you the source to run yourself for free.

PlanPricePlan FeaturesBest For
Databench (self-hosted)$0Apache-2.0 source; you run it and pay your own model providerTeams that must keep data on their own servers
Free$010 GB storage, limited usage, native integrationsTrying the hosted Alkera workspace
Plus$50/month1 TB storage, extra usage, access to all modelsIndividual analysts using the agent regularly
Pro$200/month5 TB storage, significantly more usage, dedicated support, early accessHeavy users and small data teams
EnterpriseCustomSSO/SAML, SCIM, audit logs, VPC and on-prem, data residency, SLAsLarger organisations with security reviews

Who is Databench Best For?

Use Databench if you run a data team that cannot send warehouse data to a third-party SaaS and you have someone comfortable running containers, Postgres, a workflow engine and a TLS proxy. The permission gate, sandboxing and own-key gateway are the right pieces for that buyer, and the Apache-2.0 licence lets you ship a modified copy under your own name.

Use hosted Alkera if you want the same workspace without the operations work and a monthly price you can stop at any time.

Skip Databench if you need polished dashboards for business users today, a long public track record, or a finished SOC 2 report. Skip it too if your team mostly wants a notebook with AI help rather than an agent that runs SQL against production.

Best Databench Alternatives

Hex

Hex is the mature hosted choice for SQL and Python notebooks that become shareable apps, with its own agent. Its pricing page lists a free Community plan, Professional at $36 per editor per month and Team at $75 per editor per month. It says it is audited for SOC 2 Type II. There is no self-hosted open-source edition.

Deepnote

Deepnote is a collaborative hosted notebook with AI built in. Its pricing page shows a free plan for up to 3 editors and 5 projects, and Team at $39 per editor per month billed yearly, which includes $39 of AI credits plus CPU and GPU allowances each month. Pick it if you want a polished team notebook today.

Marimo

Marimo is the open-source reactive notebook Databench builds on, Apache-2.0 with about 23,000 stars. It is free to run and gives you the same reactive Python files without the agent chats, teams or machine handling. Choose it if you want the notebook alone and are happy to bring your own assistant.

Final Verdict: Is Databench Worth It?

Databench is a serious piece of open-source software dressed in a very young repo. The permission model, sandboxing and own-key gateway show a team that has thought about what goes wrong when an agent touches production data. The Apache-2.0 licence and clear setup guide let you evaluate it for free.

We give it 6.5 out of 10. Try the self-hosted build on a laptop with a test database before anything else, and treat the hosted plans as a reasonable deal once the docs catch up with the marketing and the Pro price stops changing between pages. Revisit it in a few months, when the repo has some release history behind it.

Databench Pros & Cons

What We Like

  • Apache-2.0 licence, so you can self-host and modify it for free and pay only your model provider
  • Every agent action is classified by effect and passes a permission gate, with SQL parsed so hidden deletes are caught
  • Reactive SQL and Python notebooks that people and the agent edit together, live
  • Chats run in gVisor sandboxes on your own machines, with a gateway that uses your own Anthropic or OpenAI keys
  • Hosted Alkera has a free plan and monthly billing with no annual commitment

What Could Be Better

  • The GitHub repo was created on 6 October 2026 with one squashed commit, one contributor and no releases
  • Dashboards and a browsable lineage view are sold on the site but listed as coming soon in the docs
  • Pro is $200/month on the pricing page but $250/month in the billing docs, and plan credit allowances are not published
  • SOC 2 Type II, ISO 27001 and HIPAA are listed as in progress, and the privacy policy names no model providers

Databench FAQ

What is Databench?

Databench by Alkera is an open-source, self-hostable web workspace where people and an agent work on data together in shared chats, files and reactive SQL and Python notebooks. Alkera is the hosted commercial version with more features.

How much does Databench cost?

Databench is free under Apache-2.0 if you host it and pay your own model provider. Hosted Alkera has a $0 Free plan, Plus at $50 per month, Pro at $200 per month on the pricing page ($250 in the docs) and custom Enterprise pricing. Billing is monthly, and the terms tie each licence to one person.

Is Databench worth it?

It is worth a free trial on a test database. The permission gate and sandboxing are well designed, but the repo is one day old and some hosted features are still coming soon, so wait before relying on it for production work.

What are the best Databench alternatives?

Hex (from $36 per editor per month) and Deepnote (Team at $39 per editor per month billed yearly) are mature hosted notebook platforms with AI. Marimo is the free, open-source reactive notebook Databench builds on.

Does Databench offer a free plan?

Yes. The open-source edition is free to self-host, and hosted Alkera has a Free plan at $0 with 10 GB of storage and limited usage.

Who is Databench best for?

Data teams that want an agent working on their warehouses but need to keep data on their own infrastructure, and have someone to run containers, Postgres and a TLS proxy.

Disclosure: MakerStack is funded by featured placement fees, sponsor slots and a small number of affiliate links. Nobody paid for this review. Where any of those does apply to a review, we say so on the page. The scoring criteria are the same in every case. See our editorial policy.

This review describes Databench as we tested it on October 7, 2026. Pricing, plans and features may have changed since. We don't update reviews on request; see our update policy.