Move beyond single-turn chatbots. SwarmDroid brings a collaborative swarm of specialized autonomous AI agents to your phone—breaking complex natural-language goals into Directed Acyclic Graphs, delegating tasks, and running securely on-device with hardware-backed encryption.
The official smart contract address will be published here upon mainnet launch. Beware of fake tokens and impersonators.
The Master Agent deconstructs complex goals into a dependency task graph, dispatches parallel subtasks across specialized agents, and streams execution results.
Designed specifically for mobile constraints: low memory footprint, strict Android sandbox security, and graceful offline fallback.
Rather than hallucinating answers in a single pass, the Master Agent decomposes user requests into an executable Directed Acyclic Graph with explicit dependency management, retry bounds, and parallel execution.
API keys for Gemini, OpenAI, Claude, and OpenRouter are stored in the hardware-backed Android Keystore with AES-256-GCM encryption. Zero plain-text credentials.
When a workflow demands a novel capability not present in the pre-installed roster, the Master Agent instantiates a sandboxed agent with approved permission bounds.
Irreversible actions—such as sending emails, publishing social media posts, or modifying system settings—strictly pause for explicit user confirmation via biometric or tap dialogs.
Structured records stored locally in SQLite Room, backed by cosine-similarity vector embeddings. Inspect, export, or wipe memory at any time.
Switch effortlessly between Google Gemini, OpenAI, Anthropic Claude, OpenRouter, and local offline Ollama models without re-architecting your workflows.
Each agent operates within an isolated permission boundary, explicit tool allowlist, and independent memory scope.
Deconstructs natural-language prompts, analyzes dependencies, generates execution DAGs, assigns subtasks, and validates output.
Crawls verified web sources, extracts structured factual markdown, inspects technical documentation, and validates citations.
Drafts professional emails and message threads. Strictly holds execution for user biometric or tap approval before sending.
Writes clean Kotlin and Python, inspects syntax trees, analyzes repository architectures, and runs code inside safe sandboxes.
Composes technical posts for X and LinkedIn, suggests viral hooks and hashtags, and manages content calendars with approval gates.
Interfaces with approved Android capabilities: launching apps, sharing content, querying battery/network, and triggering text-to-speech.
Cross-references multi-source facts, resolves conflicting data points, and commits synthesized briefs to shared vector memory.
Intercepts prompt injection payloads, audits tool scopes, redacts credentials from execution logs, and enforces zero-trust boundaries.
SwarmDroid is built on the principle of least privilege. Your private keys never touch unencrypted storage, and sensitive actions never execute without your explicit consent.
AI provider keys are secured on-device via Android Keystore MasterKeys and AES-256-GCM.
Emails, social media posts, and system configurations are blocked until you tap or authenticate.
Pre-execution scanning detects instruction overriding, prompt leakage, and jailbreak strings.
"Requesting permission to dispatch draft email to engineering@swarmdroid.ai with subject: 'v1.0 Milestone Release'."
Built with Kotlin, Jetpack Compose Material 3, Room SQLite, FastAPI WebSocket streaming, and Python async execution.
@Composable
fun SwarmDashboardScreen(viewModel: SwarmViewModel) {
val taskGraph by viewModel.activeTaskGraph.collectAsState()
val swarmState by viewModel.swarmState.collectAsState()
// Real-time animated Swarm canvas visualizer
SwarmVisualizerView(
graph = taskGraph,
activeAgents = swarmState.activeAgents,
onAgentClick = { agentId -> viewModel.inspectAgent(agentId) }
)
// Human-in-the-loop approval dialog triggered on sensitive tasks
if (swarmState.pendingApproval != null) {
ApprovalDialog(
request = swarmState.pendingApproval!!,
onApprove = { viewModel.approveTask(it) },
onReject = { viewModel.rejectTask(it) }
)
}
}from backend.orchestration.swarm_orchestrator import swarm_orchestrator
from backend.security.security_core import PromptDefense
# 1. Evaluate incoming prompt against injection attacks
defense = PromptDefense.evaluate(user_prompt)
if not defense.is_safe:
raise SecurityViolation("Prompt injection detected")
# 2. Deconstruct goal into DAG & orchestrate parallel agents
result = await swarm_orchestrator.handle_user_request(
prompt="Research Android AI runtimes and draft technical summary",
user_id="user_mobile_01"
)
# 3. Stream real-time events over WebSocket to Android client
await websocket_manager.broadcast_event(result)