Skip to main content

Features

MesaLogo provides a comprehensive set of production-verified feature modules covering the full lifecycle of multi-agent collaboration.

Dual-Engine Rule System​

Natural Language Rule Engine (LLM)​

  • Rules marked as llm type, interpreted and executed by large language models
  • Supports complex semantic conditions, e.g.: "When requirements change frequently and affect core functionality, reassess the project timeline"
  • Enables non-technical users to define business rules in natural language

Logic Rule Engine (Logic)​

  • Rules marked as logic type, executed in a secure sandbox (RuleSandbox) running JavaScript
  • Sandbox limits: 5-second timeout, 50MB memory cap, isolated environment variables
  • Supports precise computation and deterministic logic, e.g.: "When defect density > 2/KLOC and test coverage < 85%, postpone release"

Rule Set Management​

  • Rules organized via RuleSet with priority ordering
  • Rule sets bind to Action Spaces and automatically apply in task instances
  • Rule trigger logging (RuleTriggerLog) tracks context, variable values, and execution results for each check

Supervisor Mechanism​

Supervisor Roles​

  • Action Spaces can configure supervisors (ActionSpaceObserver) linked to roles marked with is_observer_role=True
  • Supervisor agents are automatically created in tasks, marked as is_observer=True

Event-Driven Supervision​

  • SupervisorEventManager triggers checks at key points:
    • Immediate supervision: checks after every agent response
    • Round-based supervision: checks after each conversation round
  • Manual rule check triggering supported

Rule Checking Flow​

  • SupervisorRuleChecker builds conversation context and executes rule checks sequentially
  • Natural language rules evaluated via ModelClient calling LLM for compliance judgment
  • Logic rules executed in Node.js sandbox via RuleSandbox
  • Generates check summaries with pass/fail statistics and detailed results

Multi-Agent Conversations​

Conversation Modes​

  • Sequential mode: agents speak in predetermined order
  • Panel mode: multiple agents respond in parallel to the same message

Smart Dispatch​

  • SmartDispatchService automatically selects the most suitable agent based on message content
  • Matches agent expertise domains and analyzes conversation context

Autonomous Discussion​

  • AutonomousTask supports multiple automation types:
    • discussion: multi-round auto-discussion with configurable rounds and topics
    • conditional_stop: auto-stop when conditions are met
    • variable_trigger: triggered by variable changes
    • time_trigger: time-based triggers
    • autonomous_scheduling: self-scheduling
  • Supports streaming responses (SSE) for real-time agent reply delivery

Conversation Summarization​

  • SummaryService generates conversation summaries
  • New sub-tasks can automatically carry the previous conversation's summary as context

MCP Tool System​

Communication Protocols​

MCP Server Manager (MCPServerManager) supports four communication protocols:

  • stdio: standard I/O, suitable for local tools
  • streamable_http: StreamableHTTP protocol, suitable for remote services
  • sse: Server-Sent Events, suitable for streaming responses
  • http: simple HTTP API with OpenAPI specification support

Server Management​

  • Loads MCP server definitions from JSON configuration files
  • Supports dynamic start, stop, and reload of servers
  • Built-in variables server (variables-server) providing agents read/write access to variables
  • Tool schema caching (ToolSchemaCache) for improved performance

Capability System​

  • Capability model defines agent abilities (e.g., memory, knowledge_access, planner)
  • Capabilities linked to MCP tools: {"server1": ["tool1", "tool2"]}
  • Security level classification (1=low risk, 2=medium risk, 3=high risk)
  • Roles bind capabilities via RoleCapability

Knowledge Base System​

Built-in Knowledge Bases​

  • Vector Knowledge Base (vector): based on Milvus vector database
    • Document processing pipeline: upload → format conversion (PDF/Word/TXT/Markdown) → chunking → embedding → indexing
    • Hybrid retrieval support: 70% vector + 30% BM25 keyword, weighted fusion
    • Built-in Reranker service for result re-ranking
  • LightRAG Knowledge Base (lightrag): lightweight graph RAG
    • Supports naive/local/global/hybrid/mix query modes
    • Independent workspace isolation

Graph-Enhanced Memory​

  • GraphEnhancement configuration supports Graphiti/LightRAG/GraphRAG frameworks
  • Graphiti provides time-sensitive graph memory where older knowledge has lower weight
  • MemoryCapabilityService dynamically manages memory capabilities for all roles based on graph enhancement toggle
  • MemorySyncService handles memory data synchronization

External Knowledge Base Compatibility​

  • ExternalKnowledgeProvider integrates with external platforms:
    • Dify (DifyAdapter)
    • FastGPT (FastGPTAdapter)
    • RagFlow (RagFlowAdapter)
    • Custom (CustomAdapter)
  • Roles can bind both internal and external knowledge bases simultaneously

Environment Variable System​

Template-Instance Separation​

  • Action Space layer: ActionSpaceSharedVariable / ActionSpaceEnvironmentVariable define variable templates
  • Task Instance layer: ActionTaskEnvironmentVariable as task-level public variables visible to all agents
  • Agent layer: AgentVariable as agent-private variables with is_public visibility control

Variable Features​

  • All variables support history tracking (history JSON field) for value change auditing
  • Shared environment variables (SharedEnvironmentVariable) for cross-task sharing
  • ExternalVariableMonitor watches for external variable changes
  • Variables exposed to agents via MCP variables server for read/write access

Entity Application Marketplace​

Built-in Applications​

  • MarketApp model manages marketplace applications
  • Built-in apps include: NetLogo modeling, GIS maps, Code programming servers, etc.
  • Apps bind to Action Spaces via ActionSpaceApp

VNC Remote Desktop​

  • VNC proxy manager (VNCProxy) implements WebSocket-to-VNC TCP proxy via websockify
  • Single-port + token dynamic routing architecture supporting concurrent multi-app instance access

Parallel Laboratory​

Experiment Types​

  • Comparative experiment: generates Cartesian product of variable combinations, each creating an independent ActionTask
  • Normal task: uses fixed variable values, generating a single task instance

Experiment Flow​

  • Binds an ActionSpace template as the experiment foundation
  • Automatically clones ActionTasks with isolated variables per clone (is_experiment_clone=True)
  • Reuses existing ActionTask + AutonomousTask mechanisms for experiment execution
  • ExperimentStep records status and results for each experiment step

Experiment Configuration​

  • Supports enumerated variable definitions for experiment parameter ranges
  • Supports optimization objectives: maximize/minimize specified variables
  • Supports stop conditions: auto-terminate when expressions are satisfied

One-Click Generation​

  • OneClickGenerationService uses LLM to automatically generate roles, action spaces, rules, and tasks
  • Supports configuring a dedicated assistant generation model (assistant_generation_model)
  • Generates content based on customizable prompt templates

Users & Permissions​

Authentication​

  • Local username/password authentication (Werkzeug password hashing)
  • OAuth 2.0 social login: Google, GitHub, etc. (via authlib)
  • JWT Token authentication

Access Control​

  • UserPermissionService provides resource-level permission filtering
  • Admin role support (is_admin)
  • Multi-tenant isolation: resources controlled via created_by and is_shared fields
  • UserRoleAssignment for user role assignment

Subscriptions & Quotas​

  • SubscriptionPlan defines subscription plans and resource limits
  • Stock resource checks: tasks, agents, spaces, knowledge_bases
  • Flow resource checks: daily_conversations, monthly_tokens, storage_mb
  • UsageRecord tracks resource consumption

Model Management​

Multi-Model Support​

  • ModelConfig manages model configurations: OpenAI, Anthropic Claude, Google Gemini, domestic LLMs, etc.
  • Role-level model parameter overrides: temperature, top_p, frequency_penalty, presence_penalty, stop_sequences
  • Supports configuring default text generation and assistant generation models

External Platform Compatibility​

  • Role source supports internal and external
  • Compatible with any platform using OpenAI API format (Dify, FastGPT, Coze, etc.)