Insight Orchestra
GitHub
Multi-agent data analysis · Runs on your machine

Your data, analyzed by a team of AI agents.

Connect a data file or a database and watch specialized agents clean it, form hypotheses, debate them, and visualize what matters — then ask follow-ups in plain English on a live canvas.

Self-host in two commandsStar on GitHub
Apache 2.0·Runs locally with Docker·Your data stays in your environment
licenseApache-2.0deployDockeragents5models4
insight-orchestra · localhost:8501
A real run on a 1,000-row dataset — unedited.
/02 · the stack

A focused pipeline, built to extend.

0
Specialist agents
clean, hypothesize, debate, visualize, summarize
JanitorHypothesisDebateVizSummarizer+ your own
0
Data sources
5 file formats + 4 databases — all read-only
CSVTSVExcelJSONParquetPostgresMySQLSQLiteDuckDB
0
Model providers
swap at runtime, cost meter runs live
OpenAIAnthropicDeepSeekOllama
0%
In your environment
sandboxed code, local models, no data leaves your box
Local-firstSandboxed execBring your own keyOffline with Ollama

Everything listed above is shipping today in main

How it works

Not one model. A pipeline of specialists.

Most AI data tools throw your question at a single model. Insight Orchestra runs a pipeline of focused agents, each doing one job well — and each one shows its work.

01

Data Janitor

Cleans and profiles your dataset: types, duplicates, missing values, outliers, and bias flags.

02

Hypothesis Bot

Generates candidate insights and relationships worth investigating.

03

Debate Manager

Scores and debates the hypotheses, surfacing a confidence-ranked consensus.

04

Viz Whiz

Picks the right chart for each finding and renders it, with the code behind it.

05

Insight Summarizer

Writes the plain-English narrative you read first, plus follow-up questions worth asking next.

Then it hands you a workspace — keep asking, and the canvas keeps answering.

The canvas

Big charts. Real answers. Confidence scores.

Every insight comes with the evidence behind it — a confidence bar, a business-value bar, the chart that proves it, and the Python that produced it. No black box.

  • Top insight surfaced first, with the source numbers inline
  • Overview and Results tabs — never cramped in a chat bubble
  • Suggested follow-ups generated from what the agents found
  • Switch models mid-conversation, with a live token + cost meter
insight-orchestra · localhost:8501
Insight Orchestra workspace — conversation on the left, charts on a live canvas on the right
Features

A real analysis workspace, not a chat box.

Two-pane workspace

Conversation drives; a wide canvas visualizes. Ask on the left, watch charts build on the right.

Ask in plain English

NL → sandboxed Python → charts and answers.

Live, full-size charts

Results render big on the canvas — never cramped in a chat bubble.

Pin & compare

Side-by-side compare grids to weigh findings against each other.

Refine on the fly

"Bar chart, color by region, log scale" — iterate conversationally.

Streaming answers

Responses type out live with animated canvas transitions.

Bring your own model

Switch between OpenAI, Anthropic, DeepSeek, or local Ollama — with a live token + cost meter.

OpenAIAnthropicDeepSeekOllama

Connect anything

CSV, TSV, Excel, JSON, or Parquet files, PostgreSQL, MySQL, SQLite, or DuckDB, plus demo datasets.

CSVExcelParquetPostgresMySQLDuckDB

Persistent history

Every analysis is saved — reopen, compare, or delete past runs.

Export & share

Interactive HTML, PDF, Markdown, CSV — or a read-only share link.

Command palette

⌘K to jump anywhere — keyboard-first.

Data preview

See a live sample of your cleaned data right in the canvas before you dig in.

Sandboxed execution

Generated code runs in RestrictedPython — safe by default, inspectable by you.

Ready for a team, when you are

Flip one switch for login, role-based access, SSO via OpenID Connect, self-service API keys, and an audit log. Off by default — single-user local runs never see a login screen.

LoginRolesOIDC SSOAPI keysAudit log
Open source

Yours to run, change, and keep.

Apache 2.0, self-hosted by design. There's no paid tier to graduate to and no seat count to manage — just source you can read, run, and shape around how your team actually works.

Read the docsForkStar on GitHub
Contributors
laban254Become a contributor →
Roadmap — built in public
Open issues, vote, or send a PR. Help shape what ships next.
Apache 2.0

Use it commercially, modify it, redistribute it.

Self-hosted by default

Runs on your hardware. No accounts, no seats, no usage caps.

No vendor lock-in

Your data, your model keys, your infrastructure.

Yours to extend

Add an agent to the pipeline, or fork the whole thing.

Quickstart

Up and running in two commands.

Self-host with Docker — your data and your model spend stay yours. The setup wizard asks which LLM provider to use, writes your config, and pulls prebuilt images.

Local Ollama for fully offlineBYO OpenAI / Anthropic / DeepSeek key

Prefer a single line? curl -fsSL .../install.sh | bash clones and runs the same wizard — full command in the setup guide.

~/dev
$ git clone https://github.com/laban254/insight-orchestra.git
$ cd insight-orchestra && ./setup.sh
→ open http://localhost:8501 and try a bundled demo dataset
Local-first

Your data doesn't leave your environment.

Insight Orchestra runs locally via Docker. Use a local model (Ollama) for fully offline analysis, or plug in a hosted provider with your own key. Either way, you control where your data and your spend go.

Local-first
Bring-your-own-key
Sandboxed execution
No vendor lock-in
Use cases

Ask the questions you've been putting off.

Sales & revenue

"Which regions drive revenue, and what correlates with it?"

Product & growth

"What separates our top 25% of customers from the rest?"

Ops & quality

"Flag outliers and data-quality issues before they skew reporting."

Under the hood

Built for trust and reproducibility.

Sandboxed code execution

Generated code runs in RestrictedPython — safe by default.

Every answer shows its work

Inspect and copy the Python that produced each chart.

Confidence + business-value scoring

Each hypothesis comes ranked so you know where to look.

Built on solid infrastructure

FastAPI · Next.js · Redis · Plotly · Docker Compose.

See your data think.

Self-host it and ask your first question in minutes.

Get startedRead the docs
FAQ

Questions, answered.

Do I need to know SQL or Python?
No — ask in plain English. Power users can read and copy the generated code at any time.
Where does my data go?
It stays in your environment. Use local Ollama for fully offline runs, or plug in a hosted provider with your own key.
Which models are supported?
OpenAI, Anthropic, DeepSeek, and local Ollama — switchable at runtime, with a live token + cost meter.
What can I connect?
Any CSV, TSV, Excel, JSON, or Parquet file, or a PostgreSQL, MySQL, SQLite, or DuckDB database — plus bundled demo datasets. BigQuery is available too, but it's experimental and API-only for now (no UI).
Is it open source?
Yes — Apache 2.0. Self-host with Docker Compose in two commands.
Can I share results?
Yes — export an interactive HTML report, a PDF, a Markdown summary, or a CSV, or send a read-only share link.
Can I add teammates?
Yes, optionally. Turn on the built-in auth to get login, role-based access (admin/member/viewer), SSO via OpenID Connect, and an audit log. It's off by default for local, single-user use.