Skip to content

Tool Reference

Overwatch exposes 60+ MCP tools organized by function. Each tool uses Zod schemas for input validation and returns structured JSON.

Tool Overview

Tool Purpose Read-only
get_state Full engagement briefing from graph Yes
run_lab_preflight Aggregate lab-readiness checks Yes
run_graph_health Full graph integrity report Yes
recompute_objectives Re-evaluate objective achievement status No
get_history Full activity log Yes
export_graph Complete graph dump Yes
next_task Filtered frontier candidates for scoring No
validate_action Pre-execution sanity check No
log_action_event Record action lifecycle events No
report_finding Submit new nodes/edges to the graph No
get_evidence Retrieve full-fidelity evidence by ID or list stored evidence records Yes
parse_output Parse supported tool output into findings No
query_graph Open-ended graph exploration Yes
find_paths Shortest paths to objectives Yes
register_agent Dispatch a sub-agent task (TTL-leased) No
dispatch_agents Batch-dispatch agents from frontier No
get_agent_context Scoped subgraph for an agent Yes
update_agent Mark agent task complete/failed No
agent_heartbeat Sub-agent liveness ping (extends lease) No
dispatch_subnet_agents Dispatch one agent per scope CIDR for parallel enumeration No
dispatch_campaign_agents Dispatch agents for a campaign's grouped frontier items No
manage_campaign Create, monitor, pause, resume, or abort campaigns No
get_decision_log Per-decision timeline (frontier → completed) over the activity log Yes
explain_action "Why did the agent do X?" — full chain for an action_id Yes
get_timeline Per-node/edge "what was true at time T" view Yes
get_skill RAG search over skill library Yes
suggest_inference_rule Add a custom inference rule No
ingest_bloodhound Import BloodHound JSON collections No
check_tools Inspect installed offensive tooling Yes
track_process Register a long-running scan No
check_processes Inspect tracked process state Yes
correct_graph Transactional graph repair No
run_retrospective Post-engagement analysis Yes
open_session Create persistent interactive session (SSH, PTY, socket) No
write_session Write raw bytes to a session No
read_session Cursor-based read from session buffer Yes
send_to_session [Experimental] Write + wait + read No
list_sessions List sessions with metadata Yes
update_session Update capabilities, title, ownership No
resize_session Resize terminal dimensions No
signal_session Send signal to session process No
close_session Close and destroy a session No
update_scope Confirmation-gated runtime scope expansion/contraction No
get_system_prompt Generate dynamic agent instructions from engagement state Yes
generate_report Full pentest report with findings, narrative, evidence, remediation Yes
ingest_azurehound Import AzureHound / ROADtools JSON collections No
register_mock_service Register an operator-controlled decoy / listener / relay as a graph node No

Tool Categories

State & Health

Tools for understanding the current engagement state and verifying system health.

Scoring & Validation

The core action loop — get frontier items, validate proposed actions, log execution.

Findings & Parsing

How new information enters the graph — manual findings or deterministic parsing.

Graph Exploration

Direct graph access for creative analysis beyond the scored frontier.

Agents

Sub-agent lifecycle — register, scope, heartbeat, and track parallel work. Frontier leases (register_agent) prevent agent races; the watchdog reaps silent agents past their heartbeat_ttl_seconds.

Visibility & Introspection

"What did the agent do, and why?" — the decision log, explain_action for a single action's full chain, and get_timeline for per-node/edge "what was true at time T" queries. All read-only and derived from the activity log.

Skills & Inference

Methodology guidance and dynamic rule creation.

Ingestion

Bulk data import from external tools (BloodHound, AzureHound/ROADtools).

Utilities

System-level checks and process tracking.

Sessions

Persistent interactive sessions — SSH, local PTY, and TCP socket (reverse shell). Cursor-based I/O with ownership enforcement.

Reporting

Pentest report generation with per-finding detail, attack narrative, evidence chains, and HTML export.

Operator Infrastructure

First-class graph nodes for decoys, listeners, and relays the operator stands up — Responder, ntlmrelayx, fake LDAP, redirectors, reverse-shell catchers — so captured credentials, baited callers, and relay chains attribute back to the listener that caused them. See register_mock_service.

Analysis

Post-engagement retrospective and training data export.