datacharter

DataCharter is a federated data tool designed around local data exploration and governed agent access. Its core goal is to let users query data scattered across files and databases without moving data out of their own infrastructure, while constraining AI agent permissions within explicit contracts. The project uses the Apache-2.0 license, explicitly states that there is no paid version, and is currently in pre-release status, with V1 still under development.

[Purpose and Positioning] DataCharter describes itself as a local SQL workspace: covering files and databases, performing federated queries via DuckDB, and governed by charter.yaml. Users can use it for local data exploration, data charting, and opening data access to AI agents under controlled conditions. The official slogan emphasizes "your data is explored locally, your AI agents are kept on a leash," indicating that its core scenario is the combination of data exploration and agent permission control.

[Main Capabilities]

1. Federated queries: Can query across files as well as data sources such as Postgres, Snowflake, BigQuery, DuckDB, and Excel, with federated capabilities provided by DuckDB.
2. Contract governance: Rules are defined through charter.yaml, supporting PII default deny, row filtering, English-only policy (aggregation only), canaries, flight recorder, and access diff checks on PRs.
3. Controlled agent access: Agents can only obtain read-only, PII-masked data granted by the contract; where PII is located, agents see •••.
4. Auditing and security: Provides hash-chain auditing and SIEM JSON/OTLP output; provides datacharter redteam and GovBench (a frozen corpus of 28 attacks, rated A-F).
5. Multiple entry points: A desktop application (beta, macOS Apple Silicon or Windows, unsigned until Apple secrets exist) and a shared MCP endpoint, both sharing the same core.
6. MCP integration: Supports MCP Streamable HTTP (POST /mcp), optional OAuth 2.1; provides a Helm chart and OCI images; can wrap other MCP servers, with the command datacharter mcp --guard "npx -y some-mcp-server".
7. Optional agent integration: Supports SpaceXAI, Claude Code, Ollama (--local), or any OpenAI-compatible endpoint.

[How to Use] A workspace is just a directory containing charter.yaml, queries/, and guides/, and can be committed to version control, but secrets should not be included in it. There are two ways to enter: the desktop application (no terminal required) or a shared MCP endpoint. In MCP mode, identity definitions are stored in git rather than on an official server, and row data does not leave the user's infrastructure. The official text also mentions that it can be deployed to a machine or cluster, with no telemetry and no data plane operated by DataCharter. For the CLI reference, see docs/cli.md, and for security notes, see docs/security.md.

[Applicable Scenarios] Suitable for teams that need to uniformly query multiple files and databases locally while also wanting to open limited and compliant data access to AI agents; suitable for scenarios with requirements for PII governance, auditing, and read-only constraints; also suitable for users who want to bring existing MCP servers into the same governance system through MCP.

[Limitations and Status] The project is in pre-release status, and V1 is still under development; the desktop application is beta and unsigned; the desktop side currently only covers macOS Apple Silicon or Windows. In HTTP mode, principal and grant default to deny. The official sources do not provide specific pricing information, only stating that it is Apache-2.0 and has no paid version; information such as performance metrics, a complete list of supported data sources, and installation command details is not yet available in the given materials.