MagicHarbor.ai
Hello
Balance: $0.00
Documentation
Learn how MagicHarbor works as a unified AI platform for interacting with multiple models, comparing their performance, and building structured workflows.
Overview
MagicHarbor is a unified AI platform that allows you to work with multiple AI models in a single environment.
The platform combines:
– multi-provider model access
– a unified chat interface
– Head-to-Head model comparison
– analytics and model evaluation
– optional tool-assisted responses
– support for AI agents and automation
This enables users to move from simple AI interactions to structured, task-oriented workflows.
Core Concepts
MagicHarbor is built around several core concepts:
Model
An AI model provided by services such as OpenAI, Anthropic, Google, DeepSeek, or xAI.
Provider
The platform that hosts and serves AI models.
Chat
The primary interface for interacting with models.
Dialogue
A persistent conversation that stores messages, context, selected models, and interaction history.
Head-to-Head Compare
A structured comparison mode that allows users to send the same prompt to multiple AI models and compare their responses, cost, latency, token usage, and category-level performance.
Analytics
A system that tracks usage, cost, latency, token consumption, and model-level performance metrics.
Tools
Backend-enabled capabilities that may allow supported models to access external functions, such as online search, when the Use Tools option is enabled.
Agent
A configurable system that extends the platform from interaction to task execution.
Getting Started
1. Create an account
2. Add balance to your account
3. Choose an AI provider and model
4. Enter your prompt and start chatting
5. Optionally enable Use Tools when tool-assisted responses are needed
6. Use Head-to-Head Compare to test multiple models on the same task
You can switch models at any time and compare responses within the same interface.
How It Works
When you send aa regular chat request:
1. The request is routed to the selected AI provider
2. The selected model processes the input
3. The response is returned to MagicHarbor
4. The response is rendered in structured form
5. Analytics data such as tokens, cost, and latency is recorded
You can switch models within the same dialogue to compare outputs while preserving context.
Architecture / Execution Flow
MagicHarbor uses a multi-provider architecture.
Frontend (Wix) → API Gateway (FastAPI backend) → AI Providers
The system routes each request to the selected provider and model.
Responses are normalized and rendered in a unified interface.
For tool-assisted workflows, the backend may also connect the selected model to additional tools before returning the final response.
Simplified flow:
User Prompt
→ MagicHarbor Frontend
→ FastAPI Backend / API Gateway
→ Selected AI Provider and Model
→ Optional Backend Tools
→ Normalized Response
→ Structured UI Rendering
→ Analytics Recording
Models
MagicHarbor integrates multiple AI providers:
• OpenAI
• Anthropic (Claude)
• Google (Gemini)
• DeepSeek
• xAI (Grok)
• Stability
• Recraft
Each model differs in:
– response quality
– speed / latency
– cost per token
– reasoning ability
– coding ability
– multimodal capabilities
– tool compatibility
– context handling
MagicHarbor allows you to evaluate these differences in real usage conditions.
Key Features
– Multi-provider AI access
– Fast model switching
– Unified chat workspace
– Persistent dialogues
– Support for text and image inputs
– Structured response rendering, including code, tables, and images
– Head-to-Head model comparison
– Category winners and comparison summaries
– Token, cost, and latency analytics
– Model ranking and usage insights
– Optional Use Tools mode for backend-enabled tool access
– Foundation for AI agents and workflow automation
Use Tools Option
The Use Tools option allows supported models to use backend-enabled tools when processing a prompt.
For example, when a prompt requires current or external information, MagicHarbor may route the request through additional backend tools such as online search, depending on configuration and availability.
When Use Tools is enabled:
– the request may include tool access instructions
– the backend may call external tools or services
– the model may use retrieved information to produce a more informed response
– the final answer may combine model reasoning with information retrieved from external sources
Use Tools is designed for practical workflows where the model may need information beyond its built-in knowledge.
Important notes:
– tool availability may depend on the selected model, backend configuration, and external service availability
– tool-assisted responses may differ from standard model-only responses
– external information may affect the final output
– search-based results may depend on the quality, freshness, and availability of external sources
– in some comparison workflows, tool-assisted mode may be separated from Head-to-Head Compare to keep model comparisons consistent
Head-to-Head Compare
Head-to-Head Compare allows users to evaluate several AI models on the same prompt.
This mode is useful when you want to understand which model performs best for a specific task, such as:
– writing
– coding
– analysis
– reasoning
– summarization
– translation
– product strategy
– technical explanation
In Head-to-Head Compare, MagicHarbor can run the same prompt across selected models and display structured results.
The comparison may include:
– model responses
– response time
– token usage
– estimated cost
– clarity score
– detail score
– value score
– category winners
– overall summary
This gives users a practical way to compare models under similar conditions instead of relying only on general benchmarks or assumptions.
Head-to-Head Analytics
MagicHarbor can summarize Head-to-Head results using several comparison categories.
Typical comparison categories may include:
– Most direct response
– Most detailed response
– Clearest response
– Best value
– Overall strongest result
These categories are designed to help users understand not only which model answered, but how each model performed in practical use.
Head-to-Head results should be treated as task-specific evaluations. A model that performs best on one prompt may not be the best choice for another type of task.
Usage and Limits
Usage depends on your available balance.
Each request consumes credits based on the selected model, provider pricing, token usage, and complexity of the task.
If your balance reaches zero, access to AI services is paused until you add funds.
Different models may have different cost profiles. A more advanced or slower model may consume more credits than a smaller or faster model.
Tool-assisted requests may also differ from regular requests because external tools or additional backend processing may be involved.
Session Analytics
MagicHarbor provides real-time and dialogue-based analytics to help users understand how AI models are performing during interaction.
Analytics values are calculated either:
– in real time during the current live session
– retrieved from backend aggregation for a selected dialogue and period
Messages
Total number of messages exchanged between the user and AI models.
messages = user_messages + assistant_messages
Each user input counts as one message.
Each model response counts as one message, even if it contains multiple rendered parts, such as text, code, tables, or images.
Tokens
Total number of tokens processed by the model.
tokens = input_tokens + output_tokens
input_tokens — tokens in user prompts, context, and request input
output_tokens — tokens generated by the model in the response
Values may be displayed in compact form, for example:
7.1K = 7,100 tokens
Cost
Total estimated cost of all model calls within the selected scope.
cost = Σ (input_tokens × input_price + output_tokens × output_price)
Pricing depends on the selected model and provider.
Costs are aggregated across all requests in the session, dialogue, or selected period.
For tool-assisted requests, final cost may also reflect additional backend processing or provider-side tool behavior, depending on the implementation.
Avg Latency
Average response time of AI models.
avg_latency = total_latency_ms / number_of_responses
latency_ms = response_time − request_time
Latency may be displayed in milliseconds or seconds depending on the value.
Scope
The Scope indicator defines what data is currently displayed in the analytics panel.
Live session — calculated in real time on the frontend.
Current dialogue — aggregated backend analytics for the active dialogue.
Selected period — filtered analytics, such as 24h, 7d, 30d, or all time.
Data Source & Consistency
Analytics data is collected from actual model interactions and reflects real usage.
In live mode, values may be estimated on the frontend and may differ slightly from backend totals.
When a dialogue is selected, analytics are retrieved from backend services and represent authoritative aggregated values.
Important Notes
Analytics values are dynamic and update as new messages are sent.
Complex responses, such as text + code + tables, are treated as a single assistant message but may contain multiple UI elements.
All metrics are aggregated across all models used within the selected dialogue or time period.
Provider behavior, model availability, backend routing, and external tools may affect latency, cost, and response structure.
Model Ranking
MagicHarbor evaluates AI models based on multiple factors to help users compare performance in real usage conditions.
Ranking reflects practical interaction data rather than synthetic benchmarks.
Evaluation Criteria
Response Quality
Relevance, completeness, clarity, and usefulness of the answer.
Response Speed
How quickly the model generates output.
Cost Efficiency
How much useful output is produced relative to cost.
Stability
Consistency across multiple requests and lower error rates.
User Interaction Signals
Practical usage patterns, such as repeated use, switching behavior, and continued interaction.
How Scores Are Calculated
Each model receives a composite score on a 0–10 scale based on weighted metrics.
Score = (Quality × 0.40) + (Speed × 0.25) + (Efficiency × 0.25) + (Stability × 0.10)
Each metric is normalized to a 0–10 scale before applying weights.
Metric Details
Quality — estimated based on response usefulness, clarity, completeness, and interaction patterns.
Speed — derived from average response latency; faster responses receive higher speed scores.
Efficiency — based on the relationship between output value, token usage, and cost.
Stability — based on consistency, error rate, and successful completion of requests.
About Weighting Strategy
The weights are based on practical evaluation of how users interact with AI systems:
– Quality has the highest impact because a fast but weak answer has limited value
– Speed affects the interaction flow and user productivity
– Cost efficiency matters at scale, especially for frequent or business use
– Stability ensures reliability across repeated requests
The weighting represents a balance between answer usefulness, performance, cost, and reliability.
Why These Weights
The weighting scheme is informed by:
– real usage patterns observed in multi-model environments
– common trade-offs between quality, speed, and cost
– standard practices in performance evaluation of AI systems
– the practical need to select the right model for a specific task, not just the most expensive or most popular model
Example
Model: gpt-5.4-mini
Quality: 8.5
Speed: 9.0
Efficiency: 7.5
Stability: 8.0
Final Score = 0.40×8.5 + 0.25×9.0 + 0.25×7.5 + 0.10×8.0 = 8.325
Displayed score may be rounded in the interface.
Important Note
The evaluation model is evolving.
Scores are normalized across models within the same context and may be refined over time.
Model ranking is intended as a guidance tool, not an absolute benchmark.
The best model depends on your specific task, prompt type, required quality, acceptable latency, and cost sensitivity.
Use Cases
MagicHarbor can be used for:
– comparing AI models before choosing one for a task
– testing prompts across several providers
– evaluating cost and speed differences between models
– working with text, code, tables, and visual inputs
– using tool-assisted responses for current or external information
– creating structured research or analysis workflows
– preparing model comparison reports
– building the foundation for agent-based automation
Dialogue Statistics
Dialogue Statistics shows the key usage and performance metrics for the currently opened dialogue.
This panel gives a compact operational view of the conversation:
-
Responses — the number of model responses recorded in the current dialogue.
-
Tokens — the total number of tokens used across the dialogue.
-
Cost — the estimated total model usage cost for the dialogue.
-
Avg Latency — the average response time across recorded model outputs.
These values help the user understand how heavy, expensive, or fast a dialogue has been over time.
Important note:
Dialogue Statistics is descriptive, not judgmental. It summarizes usage and speed, but does not determine which model is “best.”
Model Ranking
Model Ranking compares models that were actually used in the current dialogue and orders them using a lightweight composite score.
The ranking is designed to provide a practical working view rather than a scientific benchmark. It helps the user see which models were more efficient or operationally attractive within that specific conversation.
The table includes:
-
Msg — number of responses from the model
-
Tokens — total token usage
-
Cost — estimated total cost
-
Avg Lat — average response latency
-
Score — a composite indicator based on usage, speed, and cost efficiency
-
Important note:
Model Ranking is contextual. It reflects performance inside this dialogue only and should be treated as an operational signal, not as a universal leaderboard.
Response Insight
Response Insight highlights notable response patterns detected in the current dialogue.
Instead of ranking every model by one single number, this panel surfaces specific comparative signals such as:
-
Fastest — the model with the lowest average latency
-
Lowest cost — the model with the lowest cost per response
-
Best detail — the model that tended to produce the most detailed responses in the dialogue
In Head-to-Head comparison workflows, this panel may also reflect higher-level comparison signals derived from direct model-to-model evaluation.
This block is intended to make interpretation faster: instead of reading the entire ranking table, the user can immediately see which model stood out on a specific dimension.
Important note:
Response Insight is a guidance layer. It highlights useful patterns, but it should be interpreted alongside the full ranking and dialogue context.
AI Agents
MagicHarbor’s long-term direction includes AI agents.
Agents extend the platform from individual AI interactions to structured task execution.
An agent may combine:
– selected models
– instructions
– memory or knowledge
– backend tools
– workflow rules
– business-specific logic
This allows MagicHarbor to support more advanced automation scenarios for teams and businesses.
The agents layer builds naturally on the same foundation used by the Command Center: multi-provider access, tool integration, analytics, and structured workflow control.
Limitations
MagicHarbor depends on external AI providers and, where enabled, external tools.
This means that:
– model availability may change
– provider behavior may vary
– latency may depend on external APIs
– costs may differ by provider and model
– tool-assisted responses may depend on external search results or service availability
– AI-generated content may be incomplete, inaccurate, outdated, or require human review
Users should review important outputs carefully, especially for legal, financial, medical, technical, or business-critical decisions.
For practical questions, see the FAQ section.
Usage and Limits
Usage depends on your available balance.
Each request consumes credits based on the selected model and complexity of the task.
If your balance reaches zero, access to AI services is paused until you add funds.
Start Using MagicHarbor.ai
Create an account and begin working with multiple AI models in one place.