Budibase Review

Overview

Budibase is an open-source operations platform for building internal apps, automations and AI agents on your own data, and its 2026 direction is agents first: the repo now bills it as AI agents, automations and apps that run your operations, model agnostic. You connect data sources, build interfaces in a visual builder, wire multi-step automations and, since the agents beta of March 2026, define agents as an LLM with instructions, tools, memory and structured outputs that can read rows, write rows and trigger automations, with per-tool run-as permissions and Slack or Microsoft Teams escalation when a human needs to approve. Model support is open rather than bundled: Anthropic, OpenAI, Google, Mistral, Groq, OpenRouter or any OpenAI-compatible endpoint, including locally hosted models. Licensing is layered: the repo's own guidelines say Budibase is GPLv3 overall, the client and component libraries are MPL 2.0, paid features live in a packages/pro folder under a Business Source License that converts to GPLv3 after four years, and apps you build are not GPL-restricted. Self-hosting free is real: one all-in-one Docker image rated at 2 cores and 6 GB of RAM, a compose cluster for production, or a Helm chart published to an OCI registry. Marketing teams reach for it on the same builds as its peers: lead-routing consoles, campaign budget trackers, content approval queues and dashboards that join ad platform data with CRM records. One behavior to plan around: automations trigger on rows written through Budibase, not on rows inserted directly into an external Postgres or MySQL. External data sources documented include PostgreSQL, MySQL, MongoDB, SQL Server, Oracle, Snowflake, Elasticsearch, DynamoDB, Firestore, S3, Google Sheets and REST. Cloud plans start at $19 monthly billed annually with metered actions, and self-hosted open source has unlimited actions and users in a single workspace. We have not run Budibase in production; this assessment is based on the documented architecture and public materials.

AI Capabilities

• AI agents with tools, memory and structured outputs (beta since March 2026)
• Model-agnostic: Anthropic, OpenAI, Google, Mistral, Groq, OpenRouter, custom OpenAI-compatible
• Automation AI steps: classify, translate, summarize, extract document data

Key Integrations

How to install

1. All-in-one Docker: docker run -d -t --name=budibase -p 10000:80 -v /local/path/data:/data --restart unless-stopped budibase/budibase:latest. The docs rate this at 2 cores and 6 GB of RAM for most uses, and the UI lands on port 10000.
2. Before production, set the environment variables the docs call out: CUSTOM_DOMAIN, INTERNAL_API_KEY, JWT_SECRET, MINIO_ACCESS_KEY and MINIO_SECRET_KEY, REDIS_PASSWORD, and COUCHDB_USER with COUCHDB_PASSWORD. If unset, the image seeds them with generated UUIDs into an .env file in the mounted volume.
3. Production compose cluster: download docker-compose.yaml, .env and litellm_config.yaml from the repo's hosting folder, then docker-compose up -d. You get CouchDB, MinIO, Nginx, the Budibase worker and the Budibase server as separate services; Linux is the recommended host, with setup scripts under hosting/scripts/linux.
4. Kubernetes: helm install --create-namespace --namespace budibase budibase oci://ghcr.io/budibase/charts/budibase. The charts moved to an OCI registry, so older instructions pointing at budibase.github.io need updating. Docs recommend nodes with at least 1 GB of memory and more for volume.
5. Agents and AI environment: set GEMINI_API_KEY for knowledge features, BBAI_LITELLM_KEY on every agent-running service for Budibase AI, and configure a pgvector connection under Settings if you want retrieval over your own documents.
6. Licensing network access: paid self-hosted tiers validate against account.budibase.app over HTTPS 443, and self-hosted Budibase AI reaches Budibase cloud IPs, so both need outbound allowances; air-gapped deployment is an enterprise add-on.

Best for

Marketing ops teams that want self-hosted internal apps plus agents that can act on that data, bringing their own LLM keys and even local models, with Slack or Teams as the approval surface.

Not for

Teams whose automations must fire on rows written by other systems, organizations that need a single OSI license (the pro folder is BSL), and buyers who cannot permit any outbound license validation or AI traffic.

Hands-on notes

Budibase builds data-first: connect a datasource, then generate screens against it, then add automations. The internal Budibase DB (CouchDB-backed) is the default when nothing external exists, and the connector list covers Postgres, MySQL, MongoDB, SQL Server, Oracle, Snowflake, Elasticsearch, DynamoDB, Firestore, Redis, S3, Google Sheets and REST, plus custom datasource plugins. One documented behavior catches teams that connect a live database: row-based automations fire on rows written through Budibase, so rows inserted externally into your Postgres or MySQL do not trigger anything. Plan a webhook or a polling step for that case.

The agents are the reason to look again if you evaluated Budibase before 2026. Each workspace gets an Agents tab, and an agent is an LLM configured with instructions, tools, memory and structured outputs. Tools are discovered from your datasources and automations and granted per operation, with a run-as setting that executes either as the requesting user or as an admin, and field-level exclusions for link and formula fields. Escalation pauses the agent and asks a human in Slack or Teams before resuming. Docs also document an Agent step inside automations, agent activity monitoring, and a testing flow for evaluating agents against scenarios before production. Access went channel-first: Slack and Microsoft Teams are the primary surfaces, in-app agent chat is deprecated, and the Discord and Telegram channels were removed in v3.44.1.

The licensing is three layers rather than one, and reading the repo's own guidelines settles most questions. The core is GPLv3 overall, the client and component libraries are MPL 2.0, and paid features live in packages/pro under a Business Source License that converts to GPLv3 four years after publication; apps you build are not GPL-restricted. Two operational consequences: paid self-hosted tiers validate their license against Budibase's account portal over HTTPS, so plan outbound access (offline licenses are an enterprise add-on), and Budibase AI on self-host calls Budibase's cloud endpoints, so SSL inspection can break it.

Agent and knowledge features have their own requirements, and they are the difference between a demo and a deployment. Knowledge sources (retrieval over your documents) currently support pgvector as the only vector store, and they require a GEMINI_API_KEY in the environment or the knowledge actions are disabled. Budibase AI on self-host needs BBAI_LITELLM_KEY configured on every agent-running service, including the automation worker. If you bring your own model through the custom provider option, an OpenAI-compatible base URL plus key is the contract, which is how Ollama-style local models connect.

We have not run Budibase in production; this assessment is based on the repository, docs.budibase.com and release notes. Calibration: releases are roughly weekly (v3.44.1 shipped September 7, 2026, the day this was written), the agents beta arrived in March 2026, and the project's own usage claims do not reconcile (docs say over 200,000 teams, the homepage says 300K), so treat scale claims as marketing. Cloud pricing meters actions, defined as agent tool calls plus automation node executions plus row saves, updates and deletes, with hard monthly limits and no overage billing, so a busy automation estate needs a plan check each quarter.

Verdict

The most interesting agent story in the open-source internal-tools class, wrapped in a layered license. A strong pick if you want agents acting over your own data and can accept the operational fine print that comes with them.

Similar Tools