AVA 4.01.6k
The open-source framework that gives AI agents the power to explore, analyze, and visualize data.
End-to-End Analytics
Everything Your Agent Needs to Analyze Data.
From raw data to actionable insights — one composable framework.
- 01
Connect
- 12,480 rows
sales.csvCSV · text - 3 tables
report.xlsxExcel · 3 sheets - 6 tables
inventory.dbSQLite · files - 12 tables
prod-mysqlMySQL · database
- 02
Understand
amount · distributionμ 486 · σ 712- regionVARCHAR4 distinct · 0% null
- categoryVARCHARtop: Electronics (31%)
- amountDECIMAL42 – 9,120 · μ 486
- dateDATEJan 2 → Sep 30
- 03
Analyze
Q:“Which region drives the most revenue?”
SELECT region, SUM(amount) AS total FROM sales WHERE YEAR(date) = 2025 GROUP BY region ORDER BY total DESC;✓ 4 rows · no hallucination · answer grounded in data
- 04
Visualize
Q:“Which region drives the most revenue?”
A:East is driving the most revenue ($45,382 — about 1.4× the next region).
Trusted Analytics
Trust the Analysis, Not Just the Answer.
Ground AI reasoning in real data, reliable execution, and measurable results.
DataBench · 1,789 questions
86.70%
answer accuracy with question-relevant dataset statistics
model: GLM 5.1
AVA Workflow · default84.46%
AVA Workflow · dataset statistics86.64%
AVA Workflow · question-relevant statisticsbest86.70%
Data-Aware Reasoning
Schema and data profiling give agents the context they need before analysis.
ava.profile() computes row counts, distributions, top values, and ranges — without a single LLM call.
Real Execution
Queries run against actual data through dedicated execution engines.
Generated SQL executes on DuckDB or your Python engine — errors are caught, corrected, and retried.
Measurable Quality
Evaluate accuracy, latency, and token efficiency through reproducible benchmarks.
Every result is scored on DataBench with published methodology, timing, and token usage.
Built for Real-World Data
Connect Your Data. Run Anywhere.
A unified foundation for files, databases, and flexible execution environments.
CSVfile
JSONfile
Parquetfile
Excelfile
Textinline · LLM extraction
SQLitedatabase
MySQLdatabase
PostgreSQLdatabase
ClickHousedatabase
CSVfile
JSONfile
Parquetfile
Excelfile
Textinline · LLM extraction
SQLitedatabase
MySQLdatabase
PostgreSQLdatabase
ClickHousedatabase
One interface. Multiple data sources.
Explore all data sourcesAVA · Trusted Execution Architecture
AI Agents & Applications
SDK · CLI · Agent Skill
AVA Runtime· Trusted Execution Boundary
Execution Contract
- Request
- Result
- Error
- Status
Execution Control
- Read-only policies
- Timeout & cancellation
- Resource limits
Observability
- Execution events
- Metrics & diagnostics
- Trace correlation
Execution Orchestration
- Session Lifecycle
- Engine Selection
- Result Handling
Source & Engine Adapters
Source Layer
- CSV · Parquet · SQL Databases
- Schema · Data Access
Engine Layer
- DuckDB · Python · Remote SQL
- Browser Execution
Node.jsBrowserCustom Adapters
Developer Experience
Bring Analytics to Your Agent.
Integrate AVA through an Agent Skill, CLI, or SDK — on your terms.
Give Your Agent a Data Analytics Skill.
Equip compatible agents with a reusable workflow for exploring, analyzing, and visualizing data — using the agent’s own reasoning, with no separate model key.
bash
# install the skills for your agents
$ npx skills add antvis/AVA
# or add the marketplace in Claude Code
$ /plugin marketplace add antvis/AVA
# then, in a new agent session:
"Use the AVA skill to analyze /absolute/path/sales.csv,
compare sales by region, and generate a chart."Open Source
Open Source. Built to Extend.
Composable building blocks for the next generation of AI-powered analytics.
Open by Default
Explore the source, inspect the implementation, and contribute to the project. MIT licensed, developed in the open.
Composable Architecture
Extend data sources, engines, and analytics capabilities to match your needs — a pluggable registry, not a closed box.
Built with AntV
Visualizations are powered by the GPT-Vis-compatible syntax from the AntV ecosystem — the same stack behind G2, G6, and more.
antvis/AVAPublic · MIT
Every capability on this page ships in the open — profile, engines, and benchmarks all have reproducible code and evaluation suites in the repository.