Your AD groups are your AI access policy.
No new permission system. pLLM reads your existing AD / Entra / Okta groups and turns membership into fine-grained, auditable policy on every model, MCP server, agent, and prompt.
Identity providers
Different groups, different gateways.
Three people hit the same URL — three completely different surfaces. Scope is derived from AD groups, not hand-maintained lists.
Policies, not snowflakes.
Six core policy primitives cover 95% of what enterprises actually need. Compose them.
Budget ceilings
Hard and soft monthly spend limits per group, per key, or per project.
ad:Interns → $50 / month, throttle at 80% Model allow-lists
Pin which models each group can use. Block frontier models from unvetted teams.
ad:All-Employees → gpt-4.1-mini, claude-4.6-sonnet MCP tool scoping
Allow-list individual tools on an MCP server. Read-only by default.
ad:Contractors → github-mcp/search_* only Data egress rules
Block PII, secrets, or IP-classified data from leaving the gateway.
All groups → no raw customer PII to external providers Rate & quota limits
Per-user, per-group, per-agent RPS and token quotas with burst controls.
ad:Engineering → 200 RPS, 10M tokens/day Audit & retention
Every request, tool call, and agent step logged with redacted payloads.
PII masked · 18-month retention · SIEM-ready # bind AD groups to capabilities
policies:
- match: ad:Engineers
allow:
models: [claude-4.6-*, gpt-5]
mcp: [github, jira, internal-rag]
agents: [release-captain, *]
budget: $2000/month
rate: 200rps
- match: ad:Finance-Ops
allow:
models: [gpt-5, claude-4.6-sonnet]
mcp: [sap]
agents: [finance-bot]
guardrails:
- presidio-pii
- sap-data-egress
- match: ad:Contractors
allow:
models: [gpt-4.1-mini]
mcp: [] # no tool access
budget: $25/month