Capability 02 · Controlled data access
Govern what the LLM sees, and what users get back.
LangGrant runs inside your premises and reads rows directly — but the public LLM only ever receives schema and metadata, never row data. Users and autonomous agents see only what their role and intent allow. Token spend is bounded before a question is even answered.
Trusted in regulated industries
Deployed where data control is a hard requirement.
LangGrant is built by Windocks, whose database delivery technology is in production at healthcare payers, insurers, pharmaceutical companies, claims processors, and global retailers — environments where data movement is constrained by law, policy, or audit.
Three layers of control
Control the model. Control the user. Control the cost.
Most AI-on-data initiatives stall in security review because the controls were retrofitted. LangGrant treats data exposure, role access, and token spend as first-class policy objects, not a config file.
01 · The public LLM sees schema, not rows
Rows never leave your premises. The public LLM only ever sees structure.
LangGrant deploys inside your network or VPC and reads rows from your databases as needed to compute answers. The public LLM, on the outside, only receives the schema slice, relationships, and metadata it needs to plan the work — the row data itself never crosses the premises boundary.
- LangGrant runs inside your network or VPC; row data stays on your premises
- The public LLM receives schema, descriptions, and relationships — never rows
- Use OpenAI, Anthropic, Google, or a locally hosted model under the same policy
- Works on production data (or a standby replica); no copy into a vector store required
02 · Role and intent based access
Column-level access aligned to the role, the intent, and the agent calling in.
A finance analyst, a system administrator, and an autonomous reporting agent ask different questions and need different visibility. LangGrant maps each caller to an explicit policy: which columns are visible, which are masked, and what can be returned.
- Roles include system administrator, DBA, token budget administrator, PII-authorized
- Intent-based filters layered on top of role for autonomous agents
- Column-level visibility and masking for PII fields
- Generate a masked database for non-PII users and test environments
| Caller | customer_name | ssn | revenue |
|---|---|---|---|
| PII-authorized analyst | visible | visible | visible |
| FP&A analyst | visible | masked | visible |
| Reporting agent | masked | blocked | visible |
| Public LLM | schema only | schema only | schema only |
03 · Token budgets
Estimate before answering. Track every call. Stop when the budget is gone.
AI bills don’t surprise teams that run LangGrant. Token usage is estimated before the model is invoked, every call is tracked against a budget, and limits are enforced. Persisted Agent Workflows reuse prior work instead of paying the LLM again — but the controls hold either way.
- Token estimates returned before a question is answered
- Per-user, per-role, and per-agent budgets
- Hard limits enforced at the policy layer, not in a reporting dashboard
- Reused Agent Workflows cost zero new LLM tokens
Source database access
Highly protected source database access.
How LangGrant connects to your source databases matters as much as what it does once connected. LangGrant supports OS-trusted connections that avoid storing SQL passwords entirely, encrypts any credentials it does have to store using the strongest algorithm the host OS provides, and runs against read-only source accounts when you want belt-and-braces protection.
OS-credentialed connections, no SQL passwords stored
The LangGrant agent can connect to source databases using OS-trusted credentials — Windows Authentication, Kerberos, integrated auth — so no SQL password is ever issued, transmitted, or stored anywhere in the LangGrant stack.
Stored credentials are encrypted with the strongest OS-native algorithm
When credentials must be stored, LangGrant’s system database keeps only the encrypted form — using the most modern algorithm the host OS makes available (for example, DPAPI on Windows, equivalent OS-native key stores elsewhere). The plaintext is never persisted.
Read-only source database accounts
LangGrant can connect to your databases through a read-only account, so the connection itself cannot mutate production data — a useful additional layer on top of role and intent policy for AI workloads.
Get started
Walk through your policy on real production data.
Bring a database with real role boundaries — PII columns, regional restrictions, agent-driven workflows. We’ll show you how LangGrant enforces each layer, end to end, in a working session.
- Map your existing roles to LangGrant policy
- Test column-level PII masking on real rows
- Set token budgets and see them enforced live