Power Apps alternative: open-source low-code platforms on your own server, with local AI

A Power Apps alternative on your own servers

For internal forms, approval workflows and data apps, open-source low-code platforms such as Appsmith, Budibase, ToolJet, NocoBase and Baserow run on servers the company controls, and a language model on a local GPU adds AI functions to those apps without a cloud AI service. Which platform fits depends on three things: its licence, whether its AI features accept a local model, and which features sit in the free edition. For the model behind the apps, a workstation with one NVIDIA RTX PRO 5000 Blackwell (48 GB) is the configuration we recommend for a mid-sized IT department.

This comparison is for IT teams that build internal apps for purchasing, HR, quality or facilities and want them, and the data in them, inside the company network. Power Apps is the starting point of the comparison; each platform below has its own model of data, pages and permissions. A low-code platform is a tool in which apps are assembled from ready components, such as tables, forms and buttons, connected to a database, with code only where needed. A local model is an AI language model that runs on your own GPU instead of a provider's cloud.

If the apps should be built by people outside IT, by describing them to an AI, our guide to internal tools built with AI without developers covers that path with Dyad.

What Power Apps covers, and what a replacement needs

Power Apps' overview, last updated on 18 November 2025, describes it as "a suite of apps, services, and connectors, as well as a data platform, that provides a rapid development environment to build custom apps for your business needs". It names two kinds of apps, canvas and model-driven, uses Dataverse as its data platform and connects to other sources such as SharePoint and SQL Server. Its overview also says that with Copilot you describe the app you want and the AI designs it.

A self-hosted replacement for typical internal apps needs:

• Data. Its own database or connectors to PostgreSQL, MySQL, your ERP's database or REST APIs.
• Forms, tables and pages that a power user can build, and code where the logic needs it.
• Workflows. Approvals, notifications and scheduled jobs. Workflow tools without an app front end, such as n8n or Dify, cover part of this with a local model.
• Sign-in and roles. Single sign-on with the company accounts, role-based access and audit logs. These are often the features that separate a free edition from a paid one.
• AI. Functions inside apps, such as summarising a request or extracting fields, and help in building apps, both with a model you run yourself.

Five self-hosted platforms compared

GitHub READMEs and licence files of Appsmith, Budibase, ToolJet, NocoBase and Baserow, and their documentation on AI datasources, agents, LLM services and AI providers, read 30 October 2026.

Appsmith describes itself as a "Platform to build admin panels, internal tools, and dashboards" that "Integrates with 25+ databases and any API", under the Apache 2.0 licence, with Docker as the recommended installation. Its own Appsmith AI datasource "reached end of life on September 30, 2026"; the documentation points to datasources for OpenAI, Anthropic and Google AI instead, and the OpenAI datasource lists only an API key as its connection setting. A local model is therefore reached through Appsmith's general API connections, as any other HTTP service.

Budibase is "licensed as GPL v3", with its client and component libraries under MPL and paid features under the Business Source License. Its agents can use a list of cloud providers or "Connect to a custom provider", which, according to the documentation updated on 6 October 2026, asks depending on the provider for an API base URL, an API key or a default model name. That is the setting for a local model server. The page does not say which editions include agents; Budibase's pricing page lists unlimited agents and SSO in its free self-hosted plan, and audit logs only in Enterprise.

ToolJet is under the AGPL-3.0. Its Community Edition includes the visual app builder, its own database, more than 80 data sources, SSO support and JavaScript and Python inside apps. AI app generation, the AI query builder, AI debugging, the agent builder, audit logs, role-based access control and Git sync are listed under ToolJet AI (Enterprise). In the free edition, an app reaches a model through a data source, such as an API connection to your model server.

NocoBase shows Apache 2.0 on GitHub, and its LICENSE file, updated on 24 February 2026, adds the NocoBase License Agreement. It permits commercial use, requires the NocoBase branding and licence notices in the interface to stay (except the main logo) and prohibits offering a public no-code, low-code or AI platform built on it. Its AI employees work inside the apps with their own role and field-level permissions, and its LLM service settings list "Ollama local models" next to cloud providers, with an optional base URL; the plugin is marked "Community Edition+". The README also names coding agents, OpenCode among them, that build and configure NocoBase apps through its command-line tool.

Baserow puts "all non-premium and non-enterprise features under the MIT License". It combines a database with an application builder that can "Create applications and portals, and publish them on your own domain". Since version 2.4, AI providers are set under Admin tools, with fields for an Ollama host and its models, and Baserow uses them for AI fields, formula suggestions, AI actions in the application builder and automations, and its Kuma assistant. The documentation does not state which edition each AI feature needs.

Connecting a local model to the apps

Each of the five reaches a model over HTTP, through a built-in AI provider setting or a general API connection, and the common model servers offer an API that follows OpenAI's format. Ollama documents /v1/chat/completions, /v1/embeddings and further endpoints; vLLM offers the same kind of server. A platform's "OpenAI-compatible" or "custom provider" field then points at the address of your model server instead of a cloud service.

Typical AI functions inside internal apps are small and frequent: classify an incoming request and route it to the right approver, extract the supplier, amount and date from a text, draft the reply to a rejected request, summarise a long comment thread for the approver. Each is a short request to the model, which is why one GPU serves many apps.

Two security points belong in the setup. Ollama's documentation says of its API key: "The client requires an API key value, but Ollama ignores it." The model server therefore needs network rules and a gateway that checks who calls it. vLLM's --api-key protects only the paths under /v1 and a few other prefixes, while routes such as /metrics, /health and /tokenize answer without the key, as vLLM's security documentation stated in October 2026. Use Ollama 0.35.0 or later because of CVE-2026-103663. Set OLLAMA_NO_CLOUD=1 so that the model server offers no cloud models or web search, as Ollama's FAQ describes. Our guide to securing a GPU server covers the rest.

You can order the workstation alone, the setup of a low-code platform (Appsmith, Budibase, ToolJet or Baserow) and the local model alone, or both together: the workstation delivered with everything installed and configured; write to us and we suggest the best solution for your tasks and budget.

AI app generation with a local model

Building apps from a description is where the editions differ most. Copilot in Power Apps is part of the vendor's cloud service, and ToolJet's AI app generation belongs to its Enterprise edition. With a local model, three routes work:

• NocoBase with a coding agent. The README states that coding agents build and configure apps through NocoBase's command-line tool and skills; OpenCode, one of the agents it names, works with a local model, as our guide to Claude Code with a local model explains.
• Baserow's Kuma assistant, which creates databases and workflows from plain language; its model is chosen under AI features in the AI provider settings. The documentation neither confirms that an Ollama model works with Kuma nor names the edition that includes it, so ask Baserow and test it.
• Code-based apps written by a coding agent on the workstation, for teams that prefer plain code to a platform. Our guide to vibe coding workstations sizes that case.

Apps built by an AI still need the checks of our guide to vibe coding security before go-live: access rights on the server, no keys in the front end, logging.

Sizing the GPU behind the apps

The platforms themselves are ordinary web applications in Docker; the GPU is for the model. Memory decides how many requests it answers at once: the model and a cache for each request must fit on the card. By our users-per-card rule, 90 per cent of the memory the driver reports, less 3 GiB per card, is what the model and the caches may use. We count the cache in 16-bit, Ollama's default: for Qwen3.8-27B (Apache 2.0) a request of 32K tokens needs about 2.1 GiB, a 128K session by an app-building agent about 8.1 GiB, as our Qwen hardware guide calculates.

Our estimates by the users-per-card rule with usable memory of 25.7, 40.0, 61.5 and 83.0 GiB, Ollama's 4-bit (18 GB) and 8-bit (30 GB) files of Qwen3.8-27B and a 16-bit cache, as listed on ollama.com on 30 October 2026; not measurements.

The 48 GB RTX PRO 5000 carries the 8-bit model, the more precise of the two Ollama versions, with five requests of 32K tokens at once or one agent session of 128K. The 72 GB version keeps fifteen short requests or four agent sessions, enough for app AI and an IT team that builds apps with an agent at the same time. These counts assume the context is set to 32K or 128K and OLLAMA_NUM_PARALLEL is raised: by default Ollama answers one request per model at a time and, on the 72 GB and 96 GB cards, reserves a 256K context. On the 96 GB card, vLLM with Qwen's FP8 checkpoint and an FP8 cache holds 47 requests of 32K by the same rule, the figure our other Qwen3.8-27B guides use. For apps used across the company, the model belongs on a rack server with RTX PRO 6000 Server Edition cards, the edition we recommend for continuous operation in a rack, where it also serves chat and document search, as our guide to an AI server for companies of 10 to 100 people describes.

StorageReview measured Qwen3.8-27B in NVIDIA's NVFP4 format with vLLM 0.27.1 in a review of 14 October 2026. With a single request, every card that ran the model landed between 17 and 22 tokens per second, from the RTX PRO 4500 Blackwell Server Edition and an L40S to an RTX PRO 6000 Blackwell Workstation Edition; with 32 requests at once, each of 512 input and 512 output tokens, the RTX PRO 6000 Workstation Edition at 600 W produced 616 output tokens per second in total. A classification that returns 50 tokens is quick at those rates; a drafted reply of 300 tokens takes about 14 to 18 seconds for one request by our arithmetic. We found no published figure for the RTX PRO 5000 with this model. Its 1,344 GB/s of memory bandwidth, against 1,792 GB/s on the RTX PRO 6000 Workstation Edition, sets the ceiling for one request: about 47 tokens per second for the 8-bit model, an upper bound by our estimate from the arithmetic in our guide to vibe coding workstations.

By our estimate (the card, a 350 W processor and 15 per cent for the rest), a workstation with one RTX PRO 5000 draws about 0.75 kW at full load, and a server with two RTX PRO 6000 Server Edition cards about 1.8 kW.

Planning the move to self-hosted apps

• Start with one process. An approval workflow with a form, a table and a notification shows within one project whether the platform suits your team.
• Check the edition matrix. SSO, role-based access and audit logs are where editions differ; ToolJet, for example, lists SSO in its Community Edition and audit logs and RBAC under Enterprise.
• Read the licence for your use. The open-source licences of all five permit internal company use; NocoBase's agreement adds the branding rule and the ban on a public platform offer. If you plan to change a platform's code, your legal counsel reads its licence first. Our article on open model licences for company use covers the model side.
• Decide who maintains what. IT runs the platform, its database, backups and updates; the department owns the content of its apps.
• Keep data rules in writing. Which app may hold personal data, who approves AI functions that read it, and how long logs are kept; your data-protection officer confirms the details. AI functions in HR apps that filter job applications, evaluate candidates, allocate tasks by personal traits or evaluate staff are listed among the high-risk uses in Annex III, point 4, of the AI Act (Regulation (EU) 2024/1689); your legal counsel confirms what applies.

Describe the first two or three apps you plan, and we propose the platform setup and the GPU for their AI functions.

FAQ