DecisionBox
Autonomously query your warehouse, validate findings, and turn them into ranked actions.
- Source repo
- decisionbox-io/decisionbox-platform
- Stars
- ★ 122
- Last updated
- 1d ago
- License
- AGPL-3.0
- Primary language
- Go
- FA score
- 44/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformOpenAI API · Claude API
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- An e-commerce analytics team wants unattended checks for funnel, retention, and margin patterns, followed by SQL-validated actions.
- Not a fit if
- Teams that want a hosted SaaS and won't run MongoDB and Qdrant themselves
- Teams whose warehouse is not BigQuery, Redshift, Snowflake, PostgreSQL, Databricks, or SQL Server
- Source review
- 44/100 · Major gaps 2 safety controls not found
What does this agent do, and when should you use it?
DecisionBox is a self-hostable data-discovery platform with a Go agent and API plus a Next.js dashboard. It connects to supported warehouses, lets an agent choose investigation areas, and has it write and execute SQL iteratively. Each finding is rechecked with an independent verification query before being surfaced as a confidence-scored insight and ranked recommendation. The dashboard exposes live run progress, SQL and reasoning details, semantic search, and citation-backed Ask Insights conversations. It can run through Docker Compose, Helm, or Terraform deployments for GCP, AWS, and Azure.
A user configures a project with a warehouse connection, domain pack, profile, LLM provider, and credentials. During a discovery run, the agent explores the available data, writes and executes SQL, iterates on results, and runs an independent verification query for each claimed finding before producing insights and prioritized recommendations. MongoDB stores application data, while Qdrant with HNSW supports insight and recommendation embeddings, semantic search, related matches, and RAG-backed Ask Insights sessions. LLM integrations implement the llm.Provider Chat method, warehouse integrations implement warehouse.Provider, and custom builds can use warehouse.RegisterMiddleware(), apiserver.RegisterGlobalMiddleware(), apiserver.Run(), or agentserver.Run().
- An e-commerce analytics team wants unattended checks for funnel, retention, and margin patterns, followed by SQL-validated actions.
- A SaaS analytics lead with event and subscription data in PostgreSQL, BigQuery, or Snowflake wants automatic discovery of activation and churn signals.
- A game operations team needs analysis tailored to match-3, idle/incremental, or casual/hyper-casual behavior such as level failure and retention.
- An enterprise using Redshift, Databricks, or SQL Server wants a traceable interface for reviewing insights, searching past findings, and asking cited questions.
- A platform team wants to deploy warehouse analytics on Kubernetes and provision GKE, EKS, or AKS infrastructure through Terraform.
How do you install or deploy this agent?
Prerequisites: Docker and Docker Compose, plus connection details for a supported data warehouse and an LLM API key. Run:
git clone https://github.com/decisionbox-io/decisionbox-platform.git
cd decisionbox-platform
docker compose up -dOpen http://localhost:3000. This starts MongoDB, Qdrant, the API, and the dashboard. For local development, start only dependencies with docker compose up -d mongodb qdrant, then run make dev-api and make dev-dashboard in separate terminals.
How do you use this agent?
Create a project at http://localhost:3000 and configure its warehouse connection, domain pack, project profile, and LLM provider. Start a discovery run; the agent explores the warehouse and executes SQL while the dashboard shows live phases, steps, and expandable queries. Review confidence- and severity-ranked insights and recommendations, provide like/dislike feedback, or save items to bookmark lists. Use Ask Insights for multi-turn, citation-backed questions across prior discoveries, and use cost estimation before a run to estimate LLM-token and warehouse-query costs.
What are this agent's strengths and limitations?
- It runs autonomous discovery rather than requiring a user prompt, with documented runs of 50–100+ SQL queries.
- Every finding is re-queried through an independent verification query, while SQL, reasoning steps, and decisions remain visible for review.
- It documents six supported warehouse families and multiple LLM routes: Claude, OpenAI, Ollama, Vertex AI, Bedrock, and Azure AI Foundry.
- Domain packs are editable, importable, and exportable from the dashboard without core-code changes; e-commerce, social, and gaming packs are included.
- It offers Docker Compose, Helm, multi-architecture images, and Terraform paths for GCP, AWS, and Azure.
- A working deployment needs both a supported warehouse connection and LLM-provider credentials; discovery cannot run without them.
- The agent executes warehouse SQL, so spending depends on model tokens, query volume, and warehouse pricing even though pre-run estimation is available.
- Documented warehouse support is limited to BigQuery, Redshift, Snowflake, PostgreSQL, Databricks, and Microsoft SQL Server; another warehouse requires a warehouse.Provider implementation.
- Production operation introduces MongoDB and Qdrant plus Docker, Kubernetes, or cloud-Terraform operational responsibilities.
- The project uses AGPL-3.0, which adopters should evaluate for their distribution and deployment obligations.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| DecisionBox This agent | 44 · Major gaps | Self-hosted serviceFree + model costs | ★ 122 | 1d ago | Go | OpenAI API · Claude API |
| Astronomer Agents: AI Agent Toolkit for Airflow Data Engineering | 46 · Major gaps | Agent plugin / skillFree + model costs | ★ 458 | 7d ago | Python | Claude Code |
| DDC CWICR Construction Cost Database | 41 · Major gaps | Self-hosted serviceFree + model costs | ★ 273 | 2mo ago | HTML | OpenAI API |
| Dex Analytics Engineering Toolkit | 84 · Good | Agent plugin / skillFree | ★ 113 | today | Python | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
Evidence shows read-only enforcement, secret encryption, non-root containers, but lacks user confirmation and rollback. Data flow transparency has logging but no implementation details. Dependency security has license checks but no vulnerability scanning. External effects have cost estimation and notifications but no explicit limits. Source attribution has CLA and contributing guide, but publisher unverified.
Self-consistency is good, README aligns with CI config, but dependency availability lacks version pinning or image digests. Failure messages have timeouts and retries but no user-friendly error prompts.
Audience is clear for data-driven teams, scenarios cover multiple industries. Capability boundaries have plugin architecture and middleware but no explicit limits. Trigger precision has selective discovery but not detailed. Environment fit supports multiple deployment methods but lacks detailed configuration.
Information architecture is clear with README, docs links, directory structure. Install notes have Docker Compose and Helm. Naming stability lacks API versioning policy. Examples and FAQ have use cases but no FAQ. Known limitations not explicitly listed. License is AGPL-3.0 but no copyright info. Versioning changelog has CHANGELOG.md but no content. Maintenance responsibility has contributing guide and CLA but no explicit maintainers.
Output usability has insights and recommendations but no format details. Marginal value has autonomous discovery and validation but no comparison. Cost benefit has cost estimation but no actual data.
Claim traceability has docs and code but no implementation details. Cross-source corroboration has CI and tests but no independent verification. Fact-inference separation not explicit.
- Not found in source: confirmation before actingTurn on (or add) a confirmation step before it acts, and try it in a sandbox or test environment before real data.
- Not found in source: rollback or recovery pathBack up first, or work on a git branch or snapshot, so its changes can be undone.
- Publisher identity unverified; assess supply chain risks carefully.
- Lack of user confirmation; autonomous SQL execution may pose risks.
- Dependency security only checks licenses, no vulnerability scanning evidence.
- Rollback missing; ensure backups.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
Release notes excerpt
Each layer's output ceiling comes from its own model Two fixes. Output ceilings and reasoning effort (#504) Per-layer models let every layer run a different model — but the output ceiling was still derived from a shared default rather than from the model actually assigned to that layer. A layer pointed at a model with a smaller ceiling could be asked for more than it could return; one pointed at a larger model was capped below what it could do. …View on GitHub ↗ -
Release notes excerpt
Per-layer models now reach the projects you already had v0.37.0 gave every LLM layer its own provider and model — but only for projects written after that release. managedai.Apply stamps llm_providers, layer_models and llm_config_version on a project write. A project last written before v0.37.0 has none of them, so the layer resolver falls back to the default row and every layer quietly runs the single legacy model. …View on GitHub ↗ -
Release notes excerpt
A provider and a model for every LLM layer A project is no longer limited to one LLM provider and one model. It can hold multiple provider entries, and choose a provider and model per layer — with per-layer reasoning, prompt-cache and output-cap settings. Every layer that calls an LLM resolves its own: exploration, analysis, recommendations, reflection, schema blurbs, questions, verification and refutation, Ask and its router and suggestions, and the rest. Why it matters …View on GitHub ↗