Core Engineering Principle: Local-First
Agent Deck is engineered from the ground up as strictly local-first developer infrastructure. We believe autonomous code generation tools require uncompromising confidentiality. Your intellectual property, proprietary abstractions, uncommitted Git diffs, token payloads, and filesystem trees must remain bounded to the physical machine executing the work.
Architectural Boundary Specification
The Agent Deck local daemon process (agentdeckd) communicates over a protected local loopback Unix domain socket (/var/run/agentdeck.sock on POSIX environments or named pipes on Windows). Under no architectural circumstance does the Agent Deck client software mirror, stream, or siphon your code repository to central Agent Deck cloud servers.
Data Collected by the Desktop App
The Agent Deck Desktop client operates as an orchestration visualizer and runtime controller. The metrics processed are maintained strictly in localized, non-persistent memory buffers during active sessions.
We locally compute agent step counters, token depletion rates, context-window saturation percentages, and OS Process Identifiers (PIDs). None of these instrumentation telemetry points leave your localhost loopback adapter.
If the desktop binary encounters a unhandled panic, an error report is prompted. It is strictly opt-in, off by default, and sanitizes all system paths, environment variables, credentials, and code snippets prior to packaging.
Telemetry & The "Unknown" Standard
Our strict anti-hallucination engineering commitment guides our instrumentation. Many proprietary developer utilities synthesize or extrapolate vague performance averages when real-time operational telemetry is hidden by underlying LLM inference providers. Agent Deck repudiates this practice.
The Truth-in-Telemetry Rule
If an autonomous agent runtime or external API does not deliver deterministic, factual byte-level operational telemetry (such as exact prompt token counts, tool execution latency, or cache hits), our interface visibly outputs "Unknown".
We do not harvest cross-user behavioral benchmarks or construct predictive models from your work patterns.
Data Collected via Hardware (The Deck)
The physical console—The Deck—is built on deterministic microcontroller firmware. It acts strictly as an input/output peripheral terminal, not an intelligent edge computer.
Receives raw pixel frames and numerical gauge vectors directly from the local agentdeckd engine over USB.
Transmits momentary mechanical key switches (Accept Diff, Reject Execution, Step Over) and dual rotary optical encoder ticks.
Website & Commerce Information
When interacting with this website or participating in hardware pre-order reservations, limited operational business data is required to process commercial commitments.
- › Order Processing: Full name, verified shipping delivery address, and electronic mail coordinates for fulfillment alerts and firmware update notices.
- › Financial Transactions: Credit card numbers and bank credentials are tokenized directly through Level-1 PCI-DSS compliant payment gateways (e.g., Stripe). Agent Deck systems never store raw card numbers.
- › Corporate Entity Notice: Formal legal entity registration details, physical corporate headquarters addresses, and foreign entity subsidiaries are presently undergoing regulatory processing and will be displayed once commercial hardware shipping commences.
Analytics, Cookies & Web Tracking
Our web presence adheres to minimal footprint standards. We do not maintain marketing tracking pixels, retargeting ad beacons, or cross-site fingerprinting matrices.
Strictly functional HTTP cookies to preserve shopping cart contents during checkout and track session state for authentication portals.
We do not sell user data to advertising syndicates, nor do we permit third-party trackers to profile visitors across our documentation.
Third-Party Agent APIs (Claude, OpenAI, Copilot)
Agent Deck does not act as an intermediate proxy for your LLM prompts. The software acts as an instrumentation dashboard over autonomous agent runtimes you execute directly.
Credential Containment Model
When linking providers such as Anthropic (Claude Code), OpenAI (Codex / GPT-4o), GitHub Copilot CLI, or Cursor daemon bridges:
"auth_storage": "OS_KEYCHAIN_SECURE_ENCLAVE",
"remote_proxy": false,
"egress_allowed": "DIRECT_TO_PROVIDER_ONLY"
}
All API keys, OAuth tokens, and model configurations remain locked inside your operating system's native keychain (macOS Keychain, Linux Secret Service API, Windows Credential Manager). They are presented directly by your machine to the provider's HTTPS endpoint. Agent Deck cloud infrastructure never touches your inference credentials.
Data Retention & Local Deletion
Because Agent Deck adheres to local-first principles, you maintain sovereign control over your developer history, agent execution logs, and cache indexes at all times.
You can permanently erase all historical agent telemetry, session journals, and cache files with a single terminal instruction or via Desktop App preferences:
For online pre-orders, transaction records are retained only as required by statutory tax obligations. You may request immediate removal of pre-order reservation inquiries prior to manufacturing batch lock.
Developer Inquiries & Audits
We welcome direct security inspections, socket verification scripts, and architectural queries from systems developers and enterprise compliance officers.