Run agents, swarms & BI on your own infrastructure
The source-available agentic AI platform you deploy yourself: visual agent builder, multi-agent swarm canvas, RAG, and a full AI-native BI suite — dashboards, alerts, and an analyst that writes the SQL. One Supabase project. One Docker command. Any model provider.
Source-available (Elastic License 2.0) · Bring your own keys · Docker, bare Node, or Kubernetes
Multi-Agent Swarms
Design agent teams on a visual canvas.
Drag agents, routers, conditions, loops, sandboxed functions, evals and human-approval gates onto a canvas, wire typed inputs to outputs, and run the whole flow live with a full trace — no orchestration code. Export to LangGraph or CrewAI when you outgrow it.
Agents & routers
Agent nodes make LLM calls with tools, memory and guardrails; router agents read the input and pick one of N downstream paths.
Conditions & loops
YES/NO condition gates branch the flow, and loop nodes retry a step until it returns DONE — real control flow, no code.
Sandboxed functions
Function nodes run sandboxed JavaScript transforms with a 2-second timeout to reshape data as it passes between agents.
Human-approval gates
Approval nodes pause the run for a person to review and release, with risk flags surfaced inline before anything ships.
LLM-as-a-judge evals
Evaluate nodes score outputs judge-style, so quality gates live inside the flow instead of a separate pipeline.
Remote agents & export
Delegate to remote A2A agents, or export any swarm to LangGraph, CrewAI or the OpenAI Agents SDK when you outgrow the canvas.
Business Intelligence
A full AI-native BI suite, self-hosted.
Not a chart library bolted on — dashboards, an analyst that writes and runs the SQL, scheduled refreshes with data alerts, and an AI-built ontology of your whole data estate. All on the same instance as your agents.
One goal in, a dashboard out
Describe the goal — the analyst plans the questions, writes the SQL, runs it, picks the charts and lays everything out. Ask follow-ups in plain English, blend in knowledge-base documents, and insert any answer as a widget.
Dashboards that act
Server-side scheduled refreshes (no browser needed) with data alerts — a rule fires once when a metric crosses its threshold, re-arms when it clears, and lands in your notification bell.
An ontology of everything
The AI maps every table, warehouse and knowledge base into a drillable knowledge graph — real sample data in, field-level relationships with quoted evidence out.
MRR
$128.4k
▲ +12.4% vs prior period
Active customers
1,842
▲ +6.1% vs prior period
Churn
1.9%
▼ −0.4 pts vs prior period
MRR trend + forecast
Plan mix
Enterprise 46%
Pro 28%
Starter 16%
Revenue by region
All › EMEA · click to drill into country
Ontology
Alert: MRR crossed $120k — notified
19 visual types
Bars to box plots, KPI targets, gauges, funnels, waterfalls, pivots, filled & bubble maps, treemaps, heatmaps — plus the AI-built ontology.
Analyst-grade interactions
Click-to-drill hierarchies, cross-filtering, date-grain switching, prior-period overlays, trend lines and σ-banded forecasts.
27 data connectors
22 databases and warehouses — Postgres, MySQL, SQL Server, Oracle, Snowflake, BigQuery, Databricks, Redshift, Synapse, Trino, Athena, ClickHouse, CockroachDB and more — plus 5 apps: Google Sheets, Stripe, Shopify, HubSpot, Salesforce. Encrypted credentials, strictly read-only.
Publish anywhere
Unguessable public links, chrome-less iframe embeds, IAM group sharing, and PDF / CSV / PNG export. Viewers get their own Ask-AI panel.
Visual data prep
Drag-and-drop joins with auto-detected keys, semantic column types, live previews — materialised as reusable datasets.
Your models, governed
Every generative feature runs on the model you pick from your own integrations — BYOK, with IAM model rules enforced server-side.
IAM & model rules
Groups, invites, per-group model allowlists, read-only resource grants.
SAML SSO
Bring your identity provider; invite-only mode for locked-down instances.
Secrets vault
Write-only secrets referenced as {{secret:NAME}} across connections.
Budgets & traces
Per-user spend caps with alerts, and a full trace for every call.
Governed by design
Numbers an agent can be trusted with.
The hard part of AI analytics isn't getting an answer — it's knowing whether to act on it. A semantic layer that refuses to compile a wrong number, and an analyst that shows every step it took to reach the right one.
The metric is the contract
“revenue is declared at the order grain, and this query joins order_lines (many per order). Summing it here would multiply it.”
A refusal you can act on beats a number you can’t defend.
The AI Analyst shows its work
Revenue moved on two axes at once, so I’ll separate region from segment before attributing the fall — a single query would confound them.
EMEA fell 8.4%; enterprise renewals account for 142% of the decline, partly offset by SMB growth (step 2).
Edit any step’s SQL and the green tick expires — a badge must not outlive what it vouched for.
Row and column security, resolved per viewer
One shared model, scoped to whoever is asking. Filters compile into the SQL itself, so they hold identically on DuckDB and every warehouse — and on embedded dashboards, where a signed token from your backend decides what each of your customers sees.
An unresolvable attribute refuses the query rather than compiling an empty filter — silent zero rows read as “there is no data”.
Metrics with a grain
Declare primary keys, join cardinality and the grain a measure lives at. The compiler proves a query is safe before it runs — chasm and fan traps refused, not warned about.
Certification that means something
Draft → certified → deprecated, with certification blocked until validation passes clean. Version history, structured diffs, restore, and a dependents view of everything that would break.
Aggregate awareness
Declare a rollup table and the compiler routes to it only when it can prove the answer is identical — then tells the reader which table answered.
Analysis you can audit
Every step's SQL, result sample and self-check verdict is stored with the answer. Re-run any step, edit it, and watch the findings mark themselves stale rather than quietly disagree.
Verification that expires
A human verdict is pinned to a fingerprint of the SQL it reviewed. Change the query and the green tick voids itself — a badge must never outlive what it vouched for.
Computed, not narrated
Driver contribution, trend slopes and median/MAD outliers are arithmetic, not prose. Too little history means no forecast at all, rather than a confident line through noise.
Deploy
Live before your coffee cools.
No separate backend to run — a free-tier Supabase project is the database, auth, and storage. The app itself is one container.
Create a Supabase project
The free tier is the entire backend — Postgres, auth, storage, vector search. One `npx supabase db push` applies the full schema.
Fill in .env
Four Supabase values, plus one optional OpenRouter key if you want instance-wide, zero-config chat. Users can always bring their own keys.
docker compose up
Public config is baked at build, secrets are read at runtime. Ship it on a $5 VPS, Fly.io, Railway, Render, or Kubernetes.
Capabilities
Everything a platform needs, batteries included
The pieces you'd otherwise glue together yourself — observability, interoperability, and the guidance to use them well.
See everything your agents do
Every tool call, token, and cost lands in a full execution trace. Set per-user budgets with email alerts before a runaway loop becomes a bill — it's your key on the line, so the meter is yours too.
Interoperable by default
Connect MCP servers as agent tools, expose swarms over an A2A endpoint, and when you outgrow the canvas, export any swarm to LangGraph, CrewAI, the OpenAI Agents SDK, or Strands — your work is never trapped here.
Your whole team on one instance
Agents, swarms, BI dashboards, and interactive notebooks all ship in one workspace — so the people you invite don't just get a login, they get a shared place to build, analyze, and ship together.
Read the documentationDeck 04 · Agentic patterns
ReAct: reason, act, observe, repeat