Home · Quick start · Editor · charter.yaml · Sources · Agent · Guides · Evals · Audit · Policies · CLI · MCP · Workspace · Desktop · About · FAQ

Connecting an AI agent straight to your database is a leap of faith: most MCP database servers hand the model whatever the connection string can see. DataCharter’s MCP server takes the opposite stance — the agent gets exactly what your data contract grants, and not one column more.

uvx datacharter mcp <workspace>
# stdio MCP server: list_sources · list_tables · describe_table · query

Works with Claude Code, Claude Desktop, Cursor, Cline, Continue, Goose — anything that speaks MCP. Setup recipes per client are on the MCP page.

What “governed” means here, concretely

• Read-only by construction. The engine refuses writes, DDL, and filesystem access at the query layer — not by prompt, by parser. A misbehaving agent can’t DROP, UPDATE, or exfiltrate a file.
• Access is declared, not assumed. A charter.yaml data contract grants access per source → table → column. PII is auto-detected and defaults to no agent access; the agent sees ••• unless you explicitly grant the field.
• Policies enforced by query analysis. One YAML line gets you clean-room-style rules — aggregates only, groups of at least 10 (k-anonymity suppression), join limits. See Policies.
• Every access is recorded. A tamper-evident, hash-chained flight recorder logs each agent query with dual attribution and masked-column detail — datacharter audit verify proves the chain, and one command exports an evidence pack.
• Canary tripwires, if you want them. Fake PII rows whose unique tokens set off a tamper-evident alarm if they ever appear in agent output — opt-in canaries plant honeytokens behind the mask; if one ever appears in agent output you get a tamper-evident alarm, and block mode withholds the response.
• Context travels with the contract. Markdown guides in guides/*.md (“revenue is net of refunds”) reach every MCP client automatically, and agent evals measure whether they actually help.

One governance layer, many sources

The same contract governs local CSV/Parquet/JSON/Excel files and Postgres, MySQL, SQLite, SQL Server, Snowflake, BigQuery, Iceberg, Delta — federated by DuckDB, so the agent can join across them in one governed query. Full list on Sources.

See what the agent sees

Run datacharter serve and flip on Agent view in the UI: it shows, column by column, exactly what comes back through the MCP tools — masked fields and all. It’s the fastest way to convince yourself (or your security team) that the leash holds.

Try it in two minutes

uvx datacharter init myws --demo
uvx datacharter mcp myws        # add to your MCP client of choice

Recipes for each client are on the MCP page, and the whole thing is Apache-2.0 on GitHub. If your agent setup needs a governance story we haven’t covered, open an issue — that’s how the policies feature got built.