CyberICS AgentForge

Model

Default model for all 7 agents, with optional per-agent overrides routed through the multi-LLM gateway. Applies to new exercises.

Default LLM backend

The gateway model every agent uses unless overridden below.

Reasoning effort

Higher effort = deeper agent reasoning, slower + costlier.

Per-agent model assignmentReset all to default

Run Adversary on a strong model and Coach/AAR on a cheaper one to control cost — or pin every agent to a sovereign model from the Deployment tab for a network-isolated, self-hosted on-prem install.

Context window override

0 = use the selected model's detected window.

Fallback models

Comma-separated provider:model to try if the default fails.

Saved ✓

Deployment

Where AgentForge runs and which Supabase backend it talks to. Sovereign mode hardens the install for network-isolated, self-hosted on-prem (Souverain) use.

Deployment mode

Cloud (managed), VPS, self-hosted Docker, or network-sovereign on-prem.

Backend endpoint

Supabase / gateway base URL. Defaults to this origin for self-host.

Sovereign mode

Forces every agent to a local/sovereign model, disables external providers, and strips outbound telemetry.

Connection

Live status of the configured backend.

checking…

Saved ✓

Providers & Keys

API credentials per provider for the gateway. In Cloud mode production keys live server-side; keys entered here are used for local/self-host gateway routing.

Use my keys for autonomous exercises

Store the keys above encrypted on your account so the AI agents can run unattended exercises on your model subscription. Off = local browser only; unattended runs then use CyberICS credits (metered by your plan).

Security: browser keys stay in this browser/app. Keys saved to your account are encrypted at rest and used only server-side to run your agents; they’re never returned to the browser. For self-host, set provider keys as edge-function secrets instead.

Saved ✓

Account

Your CyberICS identity, plan tier, and session.

–

Not signed in

—

Organization

Shown on AARs and reports.

Session

—

Access token (advanced)

Paste a fresh access_token (expires hourly). Normally handled automatically.

Two-factor authentication (TOTP)

Protect this account with an authenticator app — Microsoft Authenticator, Google Authenticator, Authy. —

Saved ✓

Exercise defaults

Pre-fill values used by New exercise and the command bar. You can still change them per exercise.

Default sector

Default difficulty

Default duration

Default language

Default adversary

Default frameworks

Comma-separated. Compliance agent maps evidence to these.

Auto-start scheduled exercises

Unattended start at the scheduled time (vs. notify the facilitator).

Saved ✓

Voice & Language

Speech input/output and the interface + live-delivery language.

Voice-to-text

Dictate into the command bar and exercise composer (Web Speech).

Dictation language

Spoken replies (TTS)

Forgewright and intel summaries read aloud.

TTS voice

Speech rate

Interface & delivery language

UI text and live translation of agent output to participants.

Saved ✓

Notifications

Outbound email for invites, scheduling, and reports.

Participant invite emails

Email join links when you invite participants.

Scheduled-start reminders

Notify participants when a scheduled exercise is about to begin.

AAR-ready notifications

Email the facilitator when an after-action report is generated.

Executive readiness digest

Auto-email your board-ready risk-reduction summary on this cadence. Saved instantly.

Digest recipient

Defaults to your login email; set a board/CISO address to send there.

From address

Sender for outbound mail. Use an enclave mailer for on-prem.

Saved ✓

Memory & Context

Organizational readiness memory — recurring gaps accumulate here and feed risk, forecasting, and Forgewright.

Organizational memory

Persist gaps and recommendations across exercises.

Readiness propagation

AAR writes gap rows so risk & forecast dashboards populate.

Retention

Days to keep memory rows (0 = keep forever).

Org memory

—

Saved ✓

MCP & Tools

Model Context Protocol servers and CISA tool integrations the agents can call during analysis.

CISA import tools (CAAT topology, Malcolm/Zeek, CSET, Crossfeed, RedEye) are wired in the Tools workspace and feed agentforge-analyze.

Saved ✓

Appearance

Theme and density for the AgentForge shell.

Theme

Accent color

Density

Theme switching affects the shell chrome; print/report views already render light for stakeholders.

Saved ✓

Safety & Governance

Data handling, audit, and content guardrails.

Usage telemetry

Anonymous product analytics. Auto-disabled in Sovereign mode.

Audit log

Record facilitator actions and agent calls for compliance evidence.

Data residency

Content guardrails

Participant messages are screened for jailbreak / prompt-injection at ingress by NeMo Guardrails when the deployment sets AF_GUARDRAIL_URL (sovereign editions do). A detection is refused; if the guardrail is unreachable the message is allowed and the outage is logged — never recorded as a pass. Deployment-level setting, not per-user.

Saved ✓

About

Build information and system health.

CyberICS AgentForge

Version

v1.0

Agents

7 (Director, Adversary, Coach, Intel, Compliance, Memory, AAR)

Edge functions

15 deployed

License

CyberICS Solutions — proprietary

IP

11 provisional patents (TS-001…011) + 2 drafted

System healthRun check

Click “Run check” to ping the gateway and database.

Sessions

Scenario library

Threat intelligence

Architecture · CAAT

Compliance frameworks

Capabilities

Settings

Model

Deployment

Providers & Keys

Account

Exercise defaults

Voice & Language

Notifications

Memory & Context

MCP & Tools

Appearance

Safety & Governance

About