Operations at conversational speed.Onboard sources, forge rules, replay history one config file at a time.

Caver Intelligence is a chat-style admin orchestrator built into the Caver platform. Operators type natural-language commands and the planner decomposes them into a sequence of tool calls against the live Caver stack: running searches, writing configs, authoring detection rules, replaying history, generating compliance reports, and more. Every step is audited as an OCSF 6003 event.

The 11 shipped operator primitives cover the full day-to-day lifecycle: onboard a new data source end-to-end (onboard_sample), convert a natural-language description to a caver.toml stanza (nl_to_config), author a CAVERN rule from a threat description (forge_rule), backtest it against 90 days of history (replay_against_history), validate configs before deployment (lint_config), enrich IOCs and CVEs inline (enrich_text), diagnose rule silence (explain_rule_silence), tune thresholds to reduce false positives (tune_threshold), deploy to a canary peer first (start_canary), diff pipeline configs side-by-side (pipeline_diff), and generate a point-in-time compliance artifact (compliance_report). Two foundation primitives (search, write_config) underpin the rest.

The planner receives the operator prompt and selects from the registered tool registry. It generates a plan (a list of steps with tool names and arguments), shows the plan to the operator for approval, then dispatches each step through the dispatch loop. Safety gates block destructive operations (config overwrites, rule deletes) without an explicit confirm. The underlying model is configurable per tenant; defaults to Claude Sonnet via the Anthropic API.

No. CAVERN detection rules are deterministic SPL/OCSF queries with no LLM inference at query time. Caver Intelligence is a separate orchestration layer that uses an LLM only for planning operator commands. The two are isolated: a prompt injection in the CAVERN query path cannot reach the Intelligence orchestrator.

No. Assistant answers are grounded in real execution. It first discovers which indexes are accessible under the caller's own permissions, generates a candidate SPL query, validates it, then runs a live bounded preview against the lake before composing an answer. Every answer is a real query you can keep, re-run, or promote to a detection. Because discovery and execution run under the caller's RBAC scope, the assistant can never see or touch more than the operator who asked.

The dispatch loop enforces three gates: (1) the operator must see and approve the plan before any tool executes; (2) write operations (nl_to_config, forge_rule, write_config) require MANAGE_INTEL capability, not just RUN_SEARCH; (3) destructive operations (rule delete, peer remove) require an explicit --force flag in the tool arguments, which the planner will not inject unless the operator explicitly requested destruction. Every tool call emits an OCSF 6003 audit event regardless of outcome.

Anthropic (Claude Sonnet/Opus), OpenAI (GPT-4o, GPT-4-turbo), Azure OpenAI, and Ollama for self-hosted deployments. Provider and model are configured per-tenant via the tenancy API, so MSSP operators can give each customer their own LLM budget and provider. Prompt caching is enabled for Anthropic and Azure OpenAI to reduce per-session cost.

Yes. The Caver MCP server exposes the same tool registry over the Model Context Protocol, so operators can drive Caver Intelligence from any MCP-compatible client (Claude.ai desktop, Cursor, VS Code Continue). The orchestrator and MCP server share the same capability model: a tool visible in Intelligence is visible in MCP and vice versa.

Caver Intelligence is included in the per-deployment commercial license. LLM inference costs are operator-supplied: you bring your own API key (or use the per-tenant LLM config to centrally manage keys). RedEye does not mark up inference costs. See /caver/pricing/ for the base license structure.

Yes. The MANAGE_INTEL capability gates access to the planning and dispatch surface. Read-only plan review (view a proposed plan without executing it) is available to any operator with RUN_SEARCH. Individual tool capabilities can be restricted further: forge_rule requires MANAGE_CAVERN, start_canary requires MANAGE_PEERS, and compliance_report requires READ_SLAM_INCIDENTS. The default admin role grants all of these.