Skip to content

operator guide ai

Jump to bottom
github-actions[bot] edited this page Jun 16, 2026 · 4 revisions

Operator AI

Path: right-side Operator AI panel on every admin page (collapsed tab on the screen edge), or /ai to open it
Requires: Active engagement for chat (panel still visible elsewhere with a prompt to select an engagement)

In-dashboard red team operator assistant. Pick Auto or a specific model from the dropdown (OpenAI, Anthropic, AWS Bedrock, Ollama, etc.). The server merges SKILLS.md and skill playbooks under .cursor/skills/reaper-red-team-operator/ (red_team_operator_skills.md, mitre_attck_skills.md, and any other *_skills.md) when REAPERC2_ROOT points at the repo (Docker: /root). Set REAPER_AI_SKILLS_FILE to use one file only. Rebuild/restart after adding skills.

Configuration

Set on the ReaperC2 process (see .env.example). Configure provider credentials plus one or more models per provider.

Shared

Variable Purpose
REAPER_AI_ENABLED Set to 0 to disable all providers
REAPER_AI_DEFAULT_MODEL auto (default) or a catalog id such as openai:gpt-4o (used when UI is Auto)
REAPER_AI_DEFAULT_PROVIDER Fallback when Auto has no REAPER_AI_DEFAULT_MODEL: first model for this provider
REAPER_AI_MAX_TOKENS Max reply tokens (default 2048)

Model catalog

Live discovery (default when the provider is configured):

Provider Discovery Env to disable
Ollama GET /api/tags (ollama list) REAPER_AI_OLLAMA_DISCOVER=0
OpenAI GET /v1/models (chat models only) REAPER_AI_OPENAI_DISCOVER=0
Anthropic GET /v1/models REAPER_AI_ANTHROPIC_DISCOVER=0
Azure AI Foundry GET /openai/v1/models (OpenAI-compatible) REAPER_AI_FOUNDRY_DISCOVER=0

Discovered models are merged with curated latest IDs (gpt-5.5, claude-fable-5, etc.) and any REAPER_AI_*_MODELS list. They appear in the Operator AI dropdown even when using REAPER_AI_MODELS. Bedrock still uses explicit model IDs (enable models in the AWS console).

Optional comma-separated extras or overrides per provider:

Variable Example
REAPER_AI_OPENAI_MODELS gpt-5.5,gpt-4.1
REAPER_AI_ANTHROPIC_MODELS claude-fable-5,claude-opus-4-7,claude-sonnet-4-6
REAPER_AI_FOUNDRY_MODELS gpt-5.5,gpt-4.1 (deployment names on Azure)
REAPER_AI_OLLAMA_MODELS llama3.2,mistral
REAPER_AI_BEDROCK_MODELS global.anthropic.claude-fable-5,us.anthropic.claude-opus-4-7,us.anthropic.claude-sonnet-4-6,amazon.nova-lite-v1:0
REAPER_AI_BEDROCK_INFERENCE_PREFIX us (optional; auto from region)

Or a unified base list (discovery still adds live models for OpenAI, Anthropic, Foundry, and Ollama):

REAPER_AI_MODELS=openai:gpt-5.5,anthropic:claude-fable-5,foundry:gpt-5.5,bedrock:anthropic.claude-fable-5,ollama:llama3.2

If discovery is off and no *_MODELS is set, built-in defaults include gpt-5.5, claude-fable-5, and anthropic.claude-fable-5 (Bedrock → global.anthropic.claude-fable-5 via Converse). Selecting a model still requires access on that provider. Legacy single-model vars (REAPER_AI_OPENAI_MODEL, etc.) still work.

OpenAI

Variable Purpose
REAPER_AI_OPENAI_API_KEY API key (required for OpenAI)
REAPER_AI_OPENAI_API_URL Base URL (default https://api.openai.com/v1)
REAPER_AI_OPENAI_DISCOVER Default 1 when an API key is set: list chat models from GET /v1/models

Legacy: REAPER_AI_API_KEY, REAPER_AI_API_URL, REAPER_AI_MODEL.

Anthropic

Variable Purpose
REAPER_AI_ANTHROPIC_API_KEY API key
REAPER_AI_ANTHROPIC_API_URL Default https://api.anthropic.com/v1
REAPER_AI_ANTHROPIC_DISCOVER Default 1 when an API key is set: list models from GET /v1/models

Azure AI Foundry

Uses the OpenAI v1-compatible endpoint on your Foundry / Azure OpenAI resource (/openai/v1/chat/completions and /openai/v1/models). On Azure, the model field is often your deployment name (from discovery or REAPER_AI_FOUNDRY_MODELS). Requests use max_completion_tokens (not max_tokens) for this provider so GPT-5.x and similar deployments work, and temperature is set to 1 (Azure rejects other values for some SKUs). Deployments that do not support chat completions on this path (e.g. some Claude SKUs) return api_not_supported — use Anthropic or AWS Bedrock in ReaperC2 for those models instead.

Variable Purpose
REAPER_AI_FOUNDRY_API_KEY API key (also reads AZURE_OPENAI_API_KEY, AZURE_AI_INFERENCE_KEY)
REAPER_AI_FOUNDRY_API_URL Resource base (also reads AZURE_OPENAI_ENDPOINT); normalized to …/openai/v1
REAPER_AI_FOUNDRY_DISCOVER Default 1 when key + URL are set
REAPER_AI_FOUNDRY_MODELS Comma-separated deployment / model names
REAPER_AI_FOUNDRY_USE_API_KEY_HEADER Set 1 for legacy non-v1 Azure endpoints that expect the api-key header

Catalog id prefix: foundry: (aliases azure, azure_foundry in REAPER_AI_DEFAULT_PROVIDER).

AWS Bedrock

Uses the Bedrock Converse API. Claude Opus/Sonnet 4.x must use regional inference profile IDs (e.g. us.anthropic.claude-opus-4-7, us.anthropic.claude-sonnet-4-6 in us-east-1), not bare anthropic.claude-* foundation IDs — otherwise Converse returns on-demand throughput isn't supported. Claude Fable 5 uses a global inference profile: global.anthropic.claude-fable-5 (ReaperC2 maps bare anthropic.claude-fable-5 to that ID). Nova and many other models still use foundation IDs (amazon.nova-lite-v1:0). ReaperC2 auto-prefixes other bare Claude 4.x IDs using REAPER_AI_BEDROCK_INFERENCE_PREFIX (default derived from region: us, eu, jp, au, global).

Reasoning models can return reasoningContent blocks (chain-of-thought) from Converse in addition to or before plain text. ReaperC2 includes reasoning text in the assistant reply so you do not see a false AWS Bedrock: empty message content when the model only populated reasoning blocks.

Variable Purpose
REAPER_AI_BEDROCK_ENABLED Set to 1 to enable
REAPER_AI_BEDROCK_REGION AWS region (falls back to AWS_REGION / AWS_DEFAULT_REGION)
REAPER_AI_BEDROCK_MODELS Comma-separated Bedrock model IDs
REAPER_AI_BEDROCK_MODEL Single default model when *_MODELS is unset
REAPER_AI_BEDROCK_API_KEY Bedrock API key from the AWS console (bearer token). Also accepts AWS_BEARER_TOKEN_BEDROCK. This is not an IAM access key and does not go in SESSION_TOKEN.
REAPER_AI_BEDROCK_ACCESS_KEY_ID IAM access key ID (AKIA…) — only if using IAM keys, not a Bedrock API key
REAPER_AI_BEDROCK_SECRET_ACCESS_KEY IAM secret access key (pair with access key ID)
REAPER_AI_BEDROCK_SESSION_TOKEN Optional STS session token when using temporary IAM credentials (three-part AKIA + secret + token). Not used for Bedrock API keys.
REAPER_AI_BEDROCK_USE_IAM Set to 1 to use the default AWS credential chain (recommended on EKS with IRSA)

Which auth should I use?

  • Bedrock API key (single string from console → Generate API key): set REAPER_AI_BEDROCK_API_KEY only.
  • IAM user access keys: set ACCESS_KEY_ID + SECRET_ACCESS_KEY; add SESSION_TOKEN only if AWS gave you temporary creds.
  • EKS / instance role: REAPER_AI_BEDROCK_USE_IAM=1, no keys.

Legacy env names AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, and AWS_SESSION_TOKEN are used when the REAPER_AI_BEDROCK_* IAM vars are empty.

EKS: attach a Bedrock-capable IAM role to the ReaperC2 service account, set REAPER_AI_BEDROCK_ENABLED=1, REAPER_AI_BEDROCK_REGION, REAPER_AI_BEDROCK_USE_IAM=1, and list models — no static keys in the Secret.

Ollama

Variable Purpose
REAPER_AI_OLLAMA_ENABLED Set to 1 to enable (no API key)
REAPER_AI_OLLAMA_DISCOVER Default 1 when Ollama is enabled: populate the UI model dropdown from GET /api/tags (live ollama list)
REAPER_AI_OLLAMA_MODELS Optional comma-separated extra models (merged with discovered tags)
REAPER_AI_OLLAMA_API_URL Default http://127.0.0.1:11434/v1

Docker Compose + Ollama on the host

Platform REAPER_AI_OLLAMA_API_URL
Mac / Windows (Docker Desktop) http://host.docker.internal:11434/v1
Linux + ollama-host profile http://172.17.0.1:11434/v1

Do not add extra_hosts: host.docker.internal:host-gateway on Mac/Windows — it points host.docker.internal at the docker bridge (172.17.0.1) and Operator AI gets EOF from the socat proxy.

Do not use OLLAMA_HOST=0.0.0.0:11434 if you want to avoid LAN access — that binds on all interfaces.

Linux only: use the Compose ollama-host profile so socat listens on the docker bridge and forwards to host 127.0.0.1:11434:

  1. COMPOSE_PROFILES=ollama-host
  2. REAPER_AI_OLLAMA_API_URL=http://172.17.0.1:11434/v1

The ollama-proxy service uses network_mode: host and bind=172.17.0.1 — not exposed on your LAN. Host CLI still uses http://127.0.0.1:11434.

If docker0 is not 172.17.0.1, adjust the bind= address in docker-compose.yml (ip -4 addr show docker0).

List only models you have pulled (ollama list); e.g. gpt-oss:latest not gpt-oss if that tag is missing.

Optional host systemd alternative: deployments/docker/ollama-docker-proxy.service.example (do not run both proxy and the compose service at once).

Restart ReaperC2 after changing env vars.

Kubernetes: use ConfigMap reaperc2-ai-config and Secret reaperc2-ai-secrets from deployments/k8s/operator-ai.yaml. See Kubernetes.

UI

  • Panel — click Operator AI on the right edge of the window to expand; » collapses it. Drag the left edge of the panel to resize (300–960px; double-click the edge to reset). Width is remembered in the browser.
  • Model — Auto or any enabled catalog entry (Provider · model name).
  • Chat history — stored in sessionStorage per engagement; survives moving between Beacons, Commands, Notes, etc. in the same tab. Clear removes it. Switching engagements loads that engagement’s thread.
  • /ai — opens the panel automatically (shortcut page).

Auto uses REAPER_AI_DEFAULT_MODEL when set to a catalog id; otherwise the first model for REAPER_AI_DEFAULT_PROVIDER; otherwise the first enabled model.

What the model sees

  1. Full SKILLS.md operator playbook (system).
  2. Engagement context — name, client, notes, beacon list, recent command output (truncated).
  3. Your chat history in the panel (persisted per engagement in the browser tab until Clear or the tab closes).

Audit

Each successful chat turn is logged as ai_chat with the user prompt and assistant reply (readable in Engagement logs / All logs; full text in JSON export). Very long messages are truncated for storage.

Cursor / external agents

Same skill at repository root SKILLS.md and .cursor/skills/reaper-red-team-operator/.

Clone this wiki locally