AGENT SWARM
Local + LAN Visual Workspace  |  Concept Proposal
Prepared for: @skalesapp
Date: 2026-03-27
@v33-kind kind. curious. clauding.
Status: Concept + Interactive Prototype
Proposal Summary

From every tool I've tested in this space, I haven't found one that delivers intelligence without complexity, a companion instead of a tool, visualization without needing to write code, or value without hype. In my professional experience, the best products are user journeys and the stories they tell. Skales has the foundation to tell that story. No one else in this landscape is close.

This proposal turns vision into value. Visually. Simply. Right now, agents, skills, tasks, memory, and schedules exist inside Skales but their relationships are invisible. Users configure them in isolation. This concept brings them into a single interactive canvas where every connection is visible, every status is live, and every component can be managed from one place.

The Local Swarm maps what's inside a single instance. The Network Swarm extends the same graph across connected peers, enabling shared skills, agent lending, and task delegation between machines. One workspace, two scopes.

Beyond the visual layer, this proposal outlines a phased roadmap toward MCP and A2A protocol alignment, agent observability, and real-time interaction. These are not theoretical. The infrastructure to support them already exists inside Skales.

This concept was designed and directed by @v33-kind, prototyped with AI assistance (Claude Code). It is not production code. It is a starting point.

Agents
4+
Built-in + custom, per-agent model
Skills
25+
Built-in + custom upload
Task Engine
6 states
Anti-loop, rate limiting
Memory
7 types
Auto-extraction, dedup
Scheduling
3 systems
Cron + Autopilot + Planner
LAN Discovery
Active
mDNS + agent-sync
The Idea: Local Swarm

A visual workspace on the Agent Swarm page. Interactive canvas for visualizing the agent ecosystem, not building workflows.

1. Which agents use which skills
2. Which schedules trigger which agents
3. Which tasks feed which memories
4. What's running, what's idle, what's offline

Node cards for each component (Agent Skill Task Memory Schedule), 4 connection ports per node, auto-routing bezier curves, click-to-highlight, drag-to-reposition, snap-to-grid, zoom, pan, stats bar, create modals.

SWOT Analysis
StrengthsWeaknesses
1. Only local-first desktop app with agent viz + automatic peer discovery
2. agent-sync protocol already aligned with A2A task format
3. Skills have structured I/O, ready for MCP wrapping
4. Zero new dependencies required for MVP
5. Execution data already captured (logs, retries, timestamps)
1. No real-time push (polling only at 15s intervals)
2. No drag-to-connect edge creation yet
3. Graph readability degrades past 50+ nodes
4. JSON file storage limits concurrent write performance
5. No built-in agent cost attribution
OpportunitiesThreats
1. MCP adoption accelerating (thousands of servers, adopted by OpenAI, Google, Anthropic)
2. A2A v1.0 under Linux Foundation enables cross-framework delegation (CrewAI, LangGraph, 150+ orgs)
3. Secure skill/agent sharing across network creates local ecosystem effects
4. Task delegation between instances enables distributed workloads
5. Companion-first positioning differentiates from tool-first competitors
1. n8n (~181K stars) has native MCP (client + server nodes)
2. Dify (~135K stars) shipped bidirectional MCP in v1.6.0
3. CrewAI (~47K stars) has both MCP and A2A natively
4. MCP/A2A standards still maturing, risk of breaking changes
5. Cloud-first competitors have larger dev teams and faster release cycles
Beyond Local - Network Sharing
The Local Swarm shows what's inside your instance. The Network Swarm shows what's available across connected peers. Same graph, bigger picture.

With agent-sync already built, the same workspace enables:

1. Shared skills across instances - one Skales has a custom skill, another on the network can use it securely
2. Agent lending - a specialized agent on one machine handles tasks delegated from another
3. Task delegation across the network - assign locally, execute on whichever instance has the right agent/skill
4. Secure capability discovery - network peers advertise what they can do without exposing data
Interactive Prototype
Working HTML mockup. Not production code. Click, drag, zoom, and explore.
Local Swarm - Interactive Prototype
Node & Edge Types
Node Types
TypeColorData SourceExample
AgentIndigo #6366f1agents/definitions/Code Assistant, Content Writer
SkillGreen #22c55eskills.jsonWeb Search, Email, Vision
TaskBlue #3b82f6tasks/Write blog post, Review PR
MemoryAmber #f59e0bmemories/preference (8), fact (15)
SchedulePink #ec4899cron/Daily standup, Weekly digest
MCP ServerTeal #14b8a6Auto-discoveredClaude Desktop, Cursor MCP
Remote AgentPurple #8b5cf6LAN scan + A2APeer Skales, CrewAI, n8n
Edge Types
FromToMeaningStyle
AgentSkillAgent can use this skillSolid, source color
AgentTaskAgent executes this taskSolid, source color
TaskMemoryTask produces memorySolid, source color
ScheduleAgentSchedule triggers agentSolid, source color
MCP ServerAgentAgent has access to toolDashed, teal
Remote AgentAgentCross-instance delegationDotted, purple
Proposed Roadmap
1.
Visual Workspace
1.1 Agent relationship mapping
1.2 Skill assignment visualization
1.3 Task execution tracking
1.4 Memory category clustering
1.5 Schedule trigger connections
2.
Agent Observability
2.1 Per-agent token usage and cost attribution
2.2 Reasoning chain traces (tool calls, context, decisions)
2.3 Trust scoring (completion rate, error frequency, retries)
2.4 Task execution timeline (Gantt-style history)
2.5 Tool call audit log with parameters and responses
3.
MCP Alignment
3.1 Expose enabled Skales skills as MCP tool endpoints
3.2 Auto-discover external MCP servers on the network
3.3 Surface MCP servers as connectable nodes in the graph
3.4 Interoperability with Claude Desktop, Cursor, and MCP-compatible agents
4.
A2A Alignment
4.1 Standardize agent-sync to A2A v1.0 Agent Card format
4.2 Cross-framework task delegation (CrewAI, LangGraph, n8n)
4.3 Remote agent visualization in Network Swarm graph
4.4 Secure capability advertisement without data exposure
5.
Real-Time
5.1 SSE push model replacing 15s polling
5.2 Drag-to-connect edge creation between ports
5.3 Minimap for large graph navigation
5.4 Ollama model hot-swap from agent nodes
5.5 Live status indicators (running, idle, error)
Competitive Analysis
Market landscape as of March 2026. Research sourced from GitHub, official docs, and industry reports.
ToolStarsAgent VizPeer DiscoveryMCPA2ALocal-FirstROI Impact
Skales (with this)Growing✓ mDNSProposedProposed✓ DesktopHigh
n8n~181K✓ NativeCommunitySelf-hostMedium
Dify~135K✓ v1.6.0PluginSelf-hostMedium
CrewAI~47K✓ Native✓ Native✗ CloudMedium
LangGraph~28K✓ Platform✓ Platform✗ CloudLow
Flowise~30K✓ NativeCommunitySelf-hostLow
Mission ControlAlphaFilesystem✓ AuditMessagingMedium

Star counts approximate as of March 2026. MCP/A2A status verified against official docs and GitHub repos. Landscape is evolving rapidly.

From every tool I've tested in this space, I haven't found one that delivers intelligence without complexity, a companion instead of a tool, visualization without needing to write code, or value without hype. In my professional experience, the best products are user journeys and the stories they tell. Skales has the foundation to tell that story. No one else in this landscape is close.
Observations
1.
Integration
1.1 Could extend the existing Swarm page as a tab, keeping discovery and local visualization together
1.2 Alternatively, a standalone route allows for a dedicated full-screen experience
2.
Protocol Alignment
2.1 Skales skills already have structured inputs/outputs, making MCP wrapping relatively low effort
2.2 agent-sync is aligned with A2A task format, standardizing opens cross-framework delegation
2.3 n8n has native MCP, Dify has MCP + A2A plugin, CrewAI has both natively since v1.9.0
2.4 Combined MCP + A2A positions Skales as a node in the emerging "Agentic Mesh"
3.
Observability
3.1 Execution logs, retry counts, and timestamps are already captured
3.2 Surfacing per-agent cost tracking and reasoning traces matches Langfuse (19K+ stars, MIT) without external dependencies
3.3 Trust scoring (task completion rate, error frequency) is derivable from existing data
4.
Network Effects
4.1 Shared skills across instances create a local skill marketplace
4.2 Agent lending enables specialization without every instance needing every model
4.3 Task delegation across network distributes workloads based on capability, not location
5.
User Experience
5.1 New users with no custom agents need onboarding state (demo data, built-in agent connections)
5.2 Memory clustering by category keeps the graph readable at scale
5.3 SSE would make status changes feel instant vs. current 15s polling
6.
Market Timing
6.1 MCP is consolidating (thousands of servers across public registries, Q1 2026). Early movers get ecosystem positioning
6.2 A2A v1.0 now under Linux Foundation with 150+ org backing (Salesforce, SAP, Atlassian, AWS, Google). Window to differentiate is narrowing
6.3 Local-first AI adoption is accelerating (Ollama, LM Studio, vLLM). Users running local models are a growing segment
Final Thoughts

Most AI products treat AI as a tool. Use it, close it, forget it.

Skales treats AI as a companion. It remembers you, plans with you, works alongside you. It runs on your machine, not someone else's server. Your data stays yours.

That's rare. That's worth building on.

@v33-kind
kind. curious. clauding.
GitHub  ·  Skales Repo