Grounding data layer for AI agents
A wrong number from an agent
doesn't look wrong.
It looks like a number. Pretensor indexes your databases and pipelines into a knowledge graph of relationships, join paths, blast radius, and drift, then serves it over MCP so agents stop guessing at production. MIT-licensed, runs on your laptop.
read-only MCP tools
12
rows leave the DB
0
engines supported
4
schema::public
kuzu · graph
agent↳ traversing identity → billing
node
orgs · identity
path
orgs → users → orders → invoices → payments
Pretensor speaks MCP. Wires into the agents and databases you already run.
ClaudeCursorContinueClineZedGooseLangGraphPostgresSnowflakeBigQueryClaudeCursorContinueClineZedGooseLangGraphPostgresSnowflakeBigQuery
04 · MCP
One tool call, structured answers.
Drop Pretensor into any MCP client. The agent calls a typed tool and gets precomputed context back: search, semantic retrieval, schema, traversal, impact, code consumers, validation, and metric compilation. Twelve read-only tools. Pretensor never lets a model improvise a query against production.
Find a join path between two tables. Building SQL joins across tables.
tool traverse →
agent
// find a join path traverse( from: billing.invoices, to: core.customers, db: prod, top_k: 3, )
pretensor
→ 3 paths · confidence 0.92 · fk-backed invoices --(order_id)→ orders orders --(customer_id)→ customers sql_hint: JOIN orders ON invoices.order_id = orders.id JOIN customers ON orders.customer_id = customers.id semantic: fk_chain_promoted cached: true
10 tools · MCP 1.0 · stdio · read-only
02 · How it works
From pip install to MCP-served, in three steps.
- STEP 01
Install Pretensor
Single Python package on PyPI. Adds the `pretensor` CLI. Pick the extra for your database engine.
terminalpip install 'pretensor[postgres]'
- STEP 02
Run the setup wizard
`pretensor init` infers your DSN from the environment, indexes it, links a code repository, and registers with Claude Code, Claude Desktop, or Cursor, each with your confirmation.
terminalpretensor init → found DATABASE_URL (postgres) → indexed prod: 342 tables · 8 clusters · 312 FKs → registered with Claude Code
- STEP 03
Serve to your agent
`--config-only` prints the `mcpServers` JSON for a client `init` doesn't auto-register. Drop it into your client's MCP settings and the IDE launches the server automatically.
terminalpretensor serve --config-only { "mcpServers": { "pretensor": { "command": "pretensor", "args": ["serve", "--state-dir", ".pretensor"] } } }
+ Does Pretensor run queries against my database?
Only at index time, and only against the catalog (information_schema, pg_catalog...). After that, the graph serves every agent call. Your rows never leave the database.
+ Can an agent run its own SQL through Pretensor?
No. There is no `run_sql` tool, on any tier, ever. Pretensor never lets a model improvise a query against production. `validate_sql` checks a statement against the graph and `compile_metric` compiles one without executing it. The agent gets a verdict, not a result set.
+ How do I know the answer my agent gave me is right?
Today Pretensor removes the guessing: the agent reads FK-backed join paths instead of inferring them, `validate_sql` catches unknown tables and broken joins before the query runs, and `detect_changes` reports schema drift since the last index. Certified metrics and per-answer trust verdicts are the Pretensor Cloud direction. Join the waitlist and we'll tell you when they open.
+ Which clients can call it?
Anything that speaks MCP: Claude Desktop, Claude Code, Cursor, Cline, Zed, Continue, Goose, and any LangGraph or custom agent that wires up the MCP transport. `pretensor init` auto-registers Claude Code, Claude Desktop, and Cursor for you, with your confirmation. The server is stdio-local, so the client launches it directly.
+ Where does the graph live?
On disk, in a local Kuzu store under .pretensor/ (override with --state-dir). No cloud roundtrip, no network call at query time.
+ What about sensitive columns and PII?
Visibility profiles allow- or block-list schemas, tables, and columns, and can sync from database GRANTs. Pretensor returns shape, never values, you decide what shape is visible.
+ Is it open source?
Yes. MIT-licensed. All twelve MCP tools, the Kuzu graph builder, the connectors, and the CLI are OSS today on PyPI as `pretensor`. Pretensor Cloud is in development. Join the waitlist to become a design partner.
+ How do you handle schema drift?
Run `pretensor reindex`. It diffs the live catalog against the last snapshot and patches only what changed. The `detect_changes` MCP tool also lets agents check drift without mutating the graph.
+ What databases do you support?
Postgres, MySQL, Snowflake, and BigQuery, each through an optional extra: `pip install 'pretensor[postgres]'` or `pip install 'pretensor[snowflake]'`. Others by request. Connectors live behind a small interface.
+ How is this different from a data catalog?
A catalog is built for humans browsing tables. Pretensor is built for agents resolving them. The graph is read-only, deterministic, and accessed through twelve typed tools. Not a UI, not a search box, not a wiki.
06 · Get access
Ground truth for your data platform.
Pretensor Cloud.
↳ Hosted, with certified metrics, trust verdicts, and cross-database resolution. Design-partner access when it opens.