HelixAgent Comprehensive Feature Reference This document provides a complete reference of all features, providers, protocols, and capabilities supported by HelixAgent.
HelixAgent supports 48 LLM providers with automatic discovery and dynamic selection based on LLMsVerifier scores.
1.1 Premium Providers (Tier 1)
Provider
Default Model
Capabilities
Authentication
Priority
Claude (Anthropic) claude-sonnet-4-5-20250929Streaming, Tools, System prompts
API Key / OAuth2
1
OpenAI gpt-4oStreaming, Tools, Response formatting
API Key
1
Google Gemini gemini-2.0-flashStreaming, Tools, Safety settings, Images
API Key
2
1.2 High-Quality Specialized (Tier 2)
Provider
Default Model
Capabilities
Authentication
Priority
DeepSeek deepseek-chatStreaming, Tools
API Key
3
Mistral mistral-large-latestStreaming, Tools, Safe prompt
API Key
3
xAI (Grok) grok-3-betaStreaming, Tools, Regional support (US/EU)
API Key
3
Qwen (Alibaba) qwen-maxStreaming, Tools
API Key / OAuth2
4
Cohere command-r-plusStreaming, Tools, Citations, RAG
API Key
4
Perplexity llama-3.1-sonar-large-128k-onlineStreaming, Online search
API Key
4
AI21 Labs jamba-1.5-largeStreaming, Tools
API Key
5
1.3 Fast Inference (Tier 3)
Provider
Default Model
Capabilities
Authentication
Priority
Groq llama-3.3-70b-versatileStreaming, Tools, Audio transcription
API Key
5
Cerebras llama-3.3-70bStreaming, Fast inference
API Key
5
1.4 Alternative Providers (Tier 4)
Provider
Default Model
Capabilities
Authentication
Priority
Fireworks AI llama-v3p1-70b-instructStreaming, Tools
API Key
6
Together AI Llama-3.3-70B-Instruct-TurboStreaming, Tools
API Key
6
Replicate meta/llama-2-70b-chatAsync prediction, Webhooks
API Key
7
Hugging Face Meta-Llama-3-8B-InstructStandard/Pro modes, Cache control
API Key
8
1.5 Cloud & Inference Providers (Tier 5)
Provider
Default Model
Capabilities
Authentication
Priority
Cloudflare @cf/meta/llama-3-8b-instructWorkers AI, Streaming
API Key
6
Codestral codestral-latestCode-specialized, Streaming
API Key
6
NVIDIA meta/llama3-70b-instructNIM, Streaming, Tools
API Key
6
Hyperbolic meta-llama/Llama-3-70b-chat-hfStreaming
API Key
7
SiliconFlow deepseek-ai/DeepSeek-V2-ChatStreaming
API Key
7
Novita meta-llama/llama-3-70b-instructStreaming
API Key
7
SambaNova Meta-Llama-3-70B-InstructStreaming, Fast inference
API Key
7
Upstage solar-proStreaming
API Key
7
Sarvam sarvam-2b-v0.5Streaming, Indian languages
API Key
8
PublicAI llama-3-70bStreaming
API Key
8
1.6 Specialized Providers (Tier 6)
Provider
Default Model
Capabilities
Authentication
Priority
Kimi (Moonshot) moonshot-v1-8kStreaming, Long context
API Key
7
KimiCode moonshot-v1-codeCode-specialized, Streaming
API Key
7
Kilo kilo-v1Streaming
API Key
7
Modal meta-llama/Llama-3-70b-chat-hfServerless, Streaming
API Key
7
Nia nia-v1Streaming
API Key
8
NLPCloud chatdolphinStreaming, Tools
API Key
8
VulaVula vulavula-chatStreaming, African languages
API Key
8
Zhipu (GLM) glm-4Streaming, Tools
API Key
7
Venice llama-3.1-405bPrivacy-focused, Streaming
API Key
7
Junie junie-v1Streaming, CLI/ACP mode
API Key / OAuth2
6
1.7 Aggregators, Local & Self-Hosted (Tier 7-8)
Provider
Default Model
Capabilities
Authentication
Priority
OpenRouter anthropic/claude-3.5-sonnet150+ models, Streaming, Tools
API Key
10
Zen (OpenCode) grok-code (free)Anonymous access, Streaming, Tools
Optional API Key
4
Ollama llama3.2Local execution, Streaming
None (local)
20
LM Studio local-modelLocal execution, OpenAI-compatible
None (local)
20
Azure OpenAI gpt-4oEnterprise, Streaming, Tools
API Key
5
Vertex AI gemini-proGoogle Cloud, Streaming
Service Account
5
HelixLLM helixllm-defaultSelf-hosted, RAG, OpenAI-compatible
API Key
3
Anthropic CU claude-sonnet-4-5-20250929Computer Use integration
API Key
5
GitHub Models gpt-4oGitHub-hosted, Streaming, Tools
GitHub Token
6
Note : Ollama and LM Studio are local-only providers (score: 5.0) - used as fallback.
Provider Authentication Methods
Method
Providers
Storage
API Key Most providers (AI21, Anthropic, Cerebras, Chutes, Cloudflare, Codestral, Cohere, DeepSeek, Fireworks, Groq, HuggingFace, Hyperbolic, Kilo, Kimi, KimiCode, Mistral, Modal, Nia, NLPCloud, Novita, NVIDIA, OpenAI, OpenRouter, Perplexity, PublicAI, Replicate, SambaNova, Sarvam, SiliconFlow, Together, Upstage, Venice, VulaVula, xAI, ZAI, Zhipu)
Environment variables
OAuth2 Claude, Qwen, Junie
CLI credential files
Service Account Vertex AI, AWS Bedrock
Service account JSON / SigV4
GitHub Token GitHub Models
GitHub PAT
Azure AD Azure OpenAI
API Key or Azure AD
Anonymous Zen (free models)
Device-ID header
None Ollama, LM Studio
Local only
Self-hosted HelixLLM
API Key (optional)
HelixAgent supports 13 embedding providers with 40+ models.
Provider
Models
Dimensions
Authentication
OpenAI text-embedding-3-small, text-embedding-3-large, text-embedding-ada-0021536, 3072
API Key
Ollama nomic-embed-text, mxbai-embed-large, all-minilm384, 768, 1024
None (local)
BGE-M3 BAAI/bge-m31024
HuggingFace API Key
Nomic nomic-ai/nomic-embed-text-v1.5768
HuggingFace API Key
CodeBERT microsoft/codebert-base768
HuggingFace API Key
Qwen3 Qwen/Qwen3-Embedding-0.6B768
HuggingFace API Key
GTE thenlper/gte-large1024
HuggingFace API Key
E5 intfloat/e5-large-v21024
HuggingFace API Key
Provider
Models
Dimensions
Authentication
Cohere embed-english-v3.0, embed-multilingual-v3.0, 4 more384-4096
API Key
Voyage AI voyage-3, voyage-code-3, voyage-finance-2, 5 more512-1536
API Key
Jina AI jina-embeddings-v3, jina-clip-v1, jina-colbert-v2, 6 more128-1024
API Key
Google Vertex AI text-embedding-005, text-multilingual-embedding-002, 3 more768
Service Account
AWS Bedrock amazon.titan-embed-text-v1/v2, cohere.embed-english-v31024-1536
AWS SigV4
3. Protocol Implementations 3.1 MCP (Model Context Protocol) Total: 79+ implementations (19 adapters + 60+ containerized servers)
MCP Adapters (External Service Integrations)
Category
Adapters
Key Tools
Cloud Storage AWS S3, Google Drive
s3_list_buckets, s3_get/put_object, gdrive_list/get/create_file
Project Management Jira, Linear, Asana
jira_get/create_issue, linear_create_issue, asana_create_task
Version Control GitLab
gitlab_list_projects, gitlab_create_merge_request
Communication Slack
slack_post_message, slack_list_channels
Design Tools Figma, Miro, SVGMaker
figma_get_file, miro_list_boards, svgmaker_create_svg
Infrastructure Docker, Kubernetes
docker_list_containers, k8s_list_pods, k8s_list_deployments
Search Brave Search
brave_web_search, brave_image_search, brave_news_search
Analytics Datadog, Sentry
datadog_get_metrics, sentry_list_issues
Database MongoDB
mongodb_query_collection, mongodb_insert_document
Knowledge Notion
notion_query_database, notion_create_page
Automation Puppeteer
puppeteer_take_screenshot, puppeteer_extract_text
AI Generation Stable Diffusion
stable_diffusion_generate_image
MCP Servers (Backend Integrations)
Category
Servers
Vector Stores Chroma, Qdrant, Weaviate
Databases PostgreSQL, SQLite, Redis
Development Git, GitHub, Filesystem
Content Fetch, Memory, Replicate
Design Figma, Miro, SVGMaker, StableDiffusion
3.2 LSP (Language Server Protocol) 10 supported language servers
Language
Server
Priority
Full LSP Support
Go gopls100
Yes
Rust rust-analyzer100
Yes
TypeScript/JS ts-server95
Yes
Python pylsp90
Yes
Python pyright85
Partial
C/C++ clangd90
Yes
Java jdtls80
Core
PHP intelephense80
Yes
Ruby solargraph75
Partial
Lua lua-language-server70
Partial
LSP Capabilities : Completion, Hover, Definition, References, Diagnostics, Rename, Code Actions, Formatting, Signature Help
3.3 ACP (Agent Communication Protocol)
Component
Purpose
ACPManager Server discovery, capability enumeration, action execution
ACPClient HTTP/WebSocket transport, JSON-RPC 2.0, retry logic
Features : Multi-transport, exponential backoff, server synchronization, diagnostics
Database
Location
Capabilities
Qdrant internal/vectordb/qdrant/Dense/sparse vectors, filtering, namespaces
Milvus internal/vectordb/milvus/Scalable vector storage, batch operations
Pinecone internal/vectordb/pinecone/Cloud-native, metadata filtering
PgVector internal/vectordb/pgvector/PostgreSQL native vectors
Location : internal/debate/, internal/services/debate_*
Aspect
Description
Purpose Multi-round debate between LLM providers with consensus voting
Participants 25 LLMs (5 positions x 5 per position)
Topologies Mesh, Star, Chain, Tree
Phases 8-phase protocol: Dehallucination → SelfEvolvement → Proposal → Critique → Review → Optimization → Adversarial → Convergence
Voting 6 methods: Weighted (MiniMax), Majority, Borda Count, Condorcet, Plurality, Unanimous
Reflexion Episodic memory, verbal reflection, retry-and-learn loop, cross-session wisdom
Adversarial Red/Blue team multi-round attack-defend cycles
Approval Gates Configurable human-in-the-loop with REST API (approve/reject/gates endpoints)
Performance Parallel execution, response caching with TTL, early termination on consensus
Persistence PostgreSQL tables (debate_sessions, debate_turns, code_versions)
Provenance Full reproducibility tracking with 14 event types, JSON export
Learning Cross-debate lesson extraction and application
Activation POST /v1/debates
Location : internal/rag/
Component
Purpose
Pipeline Orchestrates retrieval workflow
Hybrid Search Dense + sparse retrieval fusion
HyDE Hypothetical document embeddings for query expansion
Reranker Multi-stage relevance scoring
Qdrant Integration Vector storage and retrieval
Location : internal/memory/
Feature
Description
Memory Types Episodic, semantic, procedural, working
Entity Graph Entity and relationship storage
Decay Automatic memory decay over time
Session Scope Cross-session recall
Location : internal/routing/semantic/
Feature
Description
Purpose Embedding-based request routing
Coverage 96.2% test coverage
Matching Threshold-based with top-K retrieval
Location : internal/agentic/
Feature
Description
Style LangGraph-style DAG execution
Node Types Agent, Tool, Condition, Parallel, Human-in-loop, Subgraph
Checkpointing Fault-tolerant state saving
Coverage 96.5% test coverage
Location : internal/security/
Component
Purpose
Red Team 40+ attack patterns
Guardrails Output safety constraints
PII Detection Sensitive data identification
Secure Fix Agent AI-powered vulnerability remediation
Audit Logging Security event tracking
OWASP Coverage Top 10 vulnerabilities
Location : internal/structured/
Schema Type
Description
JSON Schema Strict JSON output enforcement
Regex Pattern-based constraints
Grammar Context-free grammar validation
Enum Enumeration constraints
5.8 LLM Testing Framework Location : internal/testing/llm/
Feature
Description
Style DeepEval-style evaluation
Metrics Relevance, faithfulness, hallucination
RAGAS RAG evaluation metrics
Coverage 96.2% test coverage
5.9 Self-Improvement (RLAIF) Location : internal/selfimprove/
Component
Purpose
AI Reward Model LLM-as-judge scoring
Feedback Collector Human/AI/debate feedback
Policy Optimizer Policy update generation
Constitutional AI Principle enforcement
Location : internal/llmops/
Feature
Description
Prompt Registry Semantic versioning for prompts
A/B Testing Statistical experiment framework
Continuous Eval Automated quality monitoring
Alerting Regression detection
Location : internal/benchmark/
Benchmark
Type
SWE-Bench Software engineering
HumanEval Code generation
MBPP Python programming
MMLU Multi-task knowledge
GSM8K Math reasoning
MATH Advanced math
HellaSwag Commonsense reasoning
5.12 Optimization Pipeline Location : internal/optimization/
Component
Purpose
GPTCache Semantic caching (90%+ latency reduction)
Outlines Structured output constraints
Streaming Multi-level buffering and rate limiting
LangChain Chain optimization
LlamaIndex RAG optimization
SGLang Structured generation
Guidance Template optimization
LMQL Query language optimization
Location : internal/observability/
Exporter
Type
Jaeger Distributed tracing
Zipkin Trace collection
Langfuse LLM-specific observability
Prometheus Metrics export
Location : internal/background/
Feature
Description
Task Queue PostgreSQL/In-memory queueing
Worker Pool Auto-scaling (2-10 workers)
Resource Monitor CPU, memory, disk tracking
Stuck Detector Timeout and progress detection
States pending, queued, running, completed, failed, stuck, cancelled
Location : internal/notifications/
Channel
Description
SSE Server-Sent Events
WebSocket Bidirectional real-time
Webhooks HTTP callbacks with HMAC
Polling Event storage for polling clients
Location : internal/plugins/
Feature
Description
Discovery Automatic plugin scanning
Hot Reload File change detection
Dependencies Dependency resolution
Security Sandboxing and permissions
Versioning Semantic versioning
Health Plugin health monitoring
Location : internal/skills/
Feature
Description
Categories code, debug, search, git, deploy, docs, test, review
Matching Trigger phrase, fuzzy, semantic
Tracking Usage analytics
Hot Reload YAML configuration reload
Location : internal/tools/
21 Tools : Bash, Read, Write, Edit, Glob, Grep, WebFetch, WebSearch, Git, Task, TodoWrite, AskUserQuestion, EnterPlanMode, ExitPlanMode, Skill, NotebookEdit, KillShell, TaskOutput, and more.
Location : internal/agents/
48 Agents : OpenCode, Crush, HelixCode, Kiro, Aider, ClaudeCode, Cline, CodenameGoose, DeepSeekCLI, Forge, GeminiCLI, GPTEngineer, KiloCode, MistralCode, OllamaCode, Plandex, QwenCode, AmazonQ, AgentDeck, Bridle, CheshireCat, ClaudePlugins, ClaudeSquad, Codai, Codex, CodexSkills, Conduit, Emdash, FauxPilot, GetShitDone, GitHubCopilotCLI, GitHubSpecKit, GitMCP, GPTME, MobileAgent, MultiagentCoding, Nanocoder, Noi, Octogen, OpenHands, PostgresMCP, Shai, SnowCLI, TaskWeaver, UIUXProMax, VTCode, Warp, Continue
5.20 HelixLLM Integration Location : internal/llm/providers/helixllm/, HelixLLM/ (submodule)
Feature
Description
Self-hosted LLM OpenAI-compatible API with RAG capabilities
First-class provider Registered in provider registry, participates in ensemble and debate
TLS HTTPS endpoint with configurable TLS verification
Endpoints Chat completions, embeddings, models, health check
Location : internal/services/agentic_ensemble.go
Feature
Description
Mode Classification Automatic routing: single-provider, ensemble, tool-augmented debate, agentic loop
Tool-Augmented Debate Combines debate orchestration with tool calling for grounded reasoning
Agentic Execution Loop Plan-execute-verify cycle: task decomposition → layered execution → result verification → synthesis
Task Planning Decomposes complex queries into parallel/sequential task layers
Result Verification LLM-based verification of task execution results
Location : HelixMemory/ (submodule), internal/adapters/memory/factory_helixmemory.go
Feature
Description
Unified Memory Fuses Mem0, Cognee, Letta, and Graphiti into single engine
3-Stage Fusion Collect → Dedup → Rerank pipeline with weighted scoring
12 Power Features Codebase DNA, procedural memory, mesh, temporal, debate, context window, cross-project, MCP bridge, code gen, confidence, quality loop, snapshots
Circuit Breakers Fault tolerance for each memory backend
Active by default Opt out with -tags nohelixmemory
Location : HelixSpecifier/ (submodule), internal/adapters/specifier/adapter.go
Feature
Description
Spec-Driven Development 3-pillar architecture: SpecKit + Superpowers + GSD
7-Phase SDD Constitution → Specify → Clarify → Plan → Tasks → Analyze → Implement
Adaptive Ceremony Scales ceremony based on work granularity (5 levels)
Intent Classification Signal-based request analysis for effort classification
Active by default Opt out with -tags nohelixspecifier
5.24 HTTP/3 (QUIC) with Brotli Compression
Feature
Description
Primary Transport HTTP/3 (QUIC) via quic-go/quic-go
Fallback HTTP/2 when HTTP/3 is unavailable
Compression Brotli (primary, via andybalholm/brotli) → gzip (fallback)
Scope All HTTP clients and servers prefer HTTP/3
Category
Count
LLM Providers 48
Embedding Providers 13 (6 core + 7 extended)
MCP Implementations 79+ (19 adapters + 60+ containerized servers)
LSP Language Servers 10
ACP Components 2
Vector Databases 4
Tools 21
CLI Agents 48
Extracted Modules 41
Power Features 24+ major systems
Security Attack Patterns 40+
Debate Participants 25 LLMs
Debate Voting Methods 6
Benchmarks Supported 7
Code Formatters 32+
Quick Reference: API Endpoints
Endpoint
Protocol
Description
/v1/chat/completionsOpenAI
Chat completions (ensemble)
/v1/completionsOpenAI
Text completions
/v1/embeddingsOpenAI
Vector embeddings
/v1/debatesHelixAgent
AI debate system
/v1/mcpMCP
Model Context Protocol
/v1/lspLSP
Language Server Protocol
/v1/lsp/wsLSP
LSP WebSocket
/v1/acpACP
Agent Communication Protocol
/v1/rag/*HelixAgent
RAG operations
/v1/cogneeHelixAgent
Knowledge graph (optional)
/v1/visionHelixAgent
Image analysis
/v1/tasksHelixAgent
Background tasks
/v1/monitoring/*HelixAgent
Monitoring endpoints
/v1/startup/verificationHelixAgent
Startup verification status
/v1/bigdata/healthHelixAgent
BigData health
/v1/discoveryHelixAgent
Dynamic model discovery
/v1/scoringHelixAgent
Provider scoring
/v1/verificationHelixAgent
Provider verification
/v1/healthHelixAgent
Health check
/v1/agentic/workflowsHelixAgent
Agentic workflow orchestration
/v1/planning/{hiplan,mcts,tot}HelixAgent
AI planning algorithms
/v1/llmops/{experiments,evaluate,prompts}HelixAgent
LLM operations
/v1/benchmark/{run,results}HelixAgent
Benchmarking
/v1/qa/{sessions,findings,platforms,discover}HelixAgent
QA orchestration
/v1/ensemble/{sessions,teams}HelixAgent
Ensemble management
/v1/completion/*HelixAgent
Completion endpoints
/v1/formatHelixAgent
Code formatting
/v1/formattersHelixAgent
Formatter registry
/v1/graphqlHelixAgent
GraphQL (feature-flagged, GRAPHQL_ENABLED=true)
Last updated: 2026-04-06