Configuration of the AIHost settings is done in the provided AIHost/appsettings.json. The Standard ASP.NET Core JSON configuration is consumed by SystemWeaver.AIHost.exe. You can override per host via the appsettings.Local.json (gitignored) or environment variables. Keys that are prefixed with // in the shipped files are documentation comments, not settings. This article describes how to configure the settings. 

Prerequisites

• The SWExtension.AIExtension folder is located in your Client's swExplorerExtensions directory

• Example 

• Logging
• Aife (wire plane)
• Persistence (session store)
• Mcp (Model Context Protocol servers)
  • Mcp:Servers[] entry schema
• Chat
• Llm (router-level guards)
  • Llm:Admission (per-tenant rate-limit + concurrency; AIHost is single-tenant today)
  • Llm:Limits (input-payload guard)
  • Llm:CircuitBreaker 
  • Llm:Routing
  • Llm:AcceptedCredentialSources
• ContextCompression (proactive history compression)
  • How the parameters interact
    • Reading the knobs against the graph:
• LLMRequestSizing (input-estimate buffering)
• What's Next? 

The values shown in the below appsettings.json example are the shipped defaults with the package.

Example 

{
  "Logging": {
    "LogLevel": {
      "Default": "Information",
      "Microsoft.AspNetCore": "Warning"
    },
    "File": {
      "Enabled": true,
      "Directory": "%LOCALAPPDATA%\\SystemWeaver.AIHost\\logs",
      "RollingInterval": "Day",
      "RetainedFileCountLimit": 14
    }
  },
  "Aife": {
    "Port": 5001,
    "Llm": {
      "MaxOutputTokens": 32768,
      "MaxToolCalls": 50
    }
  },
  "Persistence": {
    "Provider": "InMemory"
  },
  "Mcp": {
    "Enabled": true,
    "Servers": [],
    "StreamableHttp": {
      "AllowPlaintextHttp": false
    }
  },
  "Chat": {
    "Tools": {
      "AllowMCPDiscoveredTools": false
    },
    "Logging": {
      "DisableLogRedaction": false
    }
  },
  "Llm": {
    "Admission": {
      "PerTenantRequestsPerSecond": 100,
      "PerTenantBurst": 200,
      "MaxConcurrentInFlight": 32,
      "DefaultTenantKey": "default"
    },
    "Limits": {
      "MaxInputBytes": 1048576,
      "MaxInputTokensEstimate": 250000
    },
    "CircuitBreaker": {
      "FailureThreshold": 5,
      "ResetTimeout": "00:00:30",
      "Enabled": true
    },
    "Routing": {
      "ReservedSafetyTokens": 512
    },
    "AcceptedCredentialSources": "Runtime"
  },
  "ContextCompression": {
    "Enabled": true,
    "UseActiveModelWindow": true,
    "CompressionTriggerRatio": 0.75,
    "ReservedResponseTokens": 4000,
    "MaxContextTokens": 200000,
    "RecentMessageCount": 10,
    "MaxMessageCount": 50,
    "TargetCompressionRatio": 0.3,
    "MaxSummaryTokens": 1000,
    "CompressToolCalls": true
  },
  "LLMRequestSizing": {
    "InputTokenBufferRatio": 0.15,
    "MinBufferTokens": 256
  }
}

An example Mcp:Servers[] entry (omitted above because the shipped array is empty):

{
  "Mcp": {
    "Enabled": true,
    "Servers": [
      {
        "Name": "sysw-mcp",
        "Transport": "Stdio",
        "Enabled": true,
        "Command": "C:\\non-install\\sysw-mcp-host\\sysw-mcp.exe",
        "Arguments": [],
        "Environment": {}
      },
      {
        "Name": "remote-tools",
        "Transport": "StreamableHttp",
        "Enabled": true,
        "Endpoint": "https://localhost:7099/mcp"
      }
    ],
    "StreamableHttp": { "AllowPlaintextHttp": false }
  }
}

The per-section reference below documents each key, its default, and valid range.

Logging

Aife (wire plane)

The AIFE endpoint is https://localhost:<Port>/ with mandatory client-cert mTLS; loopback bind covers IPv4 + IPv6. There is no Host key (loopback is hard-wired) and the named-pipe trust handshake is automatic (no key).

Persistence (session store)

Mcp (Model Context Protocol servers)

Mcp:Servers[] entry schema

Note: MCP-discovered tools are additionally gated by Chat:Tools:AllowMCPDiscoveredTools (Chat).

Example:  Stdio transport (launch a local MCP server process; no TLS guard applies):

  "Mcp": {
    "Enabled": true,
    "Servers": [
      {
        "Name": "sysw-mcp",
        "Transport": "Stdio",
        "Enabled": true,
        "Command": "C:\\non-install\\sysw-mcp-host\\sysw-mcp.exe",
        "Arguments": [ "--mode", "stdio" ],
        "Environment": {
          "SYSW_MCP_LOGLEVEL": "Information"
        }
      }
    ]
  }
}

Example:  StreamableHttp transport (connect to an HTTP MCP server; subject to the TLS guard):

{
  "Mcp": {
    "Enabled": true,
    "Servers": [
      {
        "Name": "remote-tools",
        "Transport": "StreamableHttp",
        "Enabled": true,
        "Endpoint": "https://localhost:7099/mcp"
      }
    ],
    "StreamableHttp": {
      "AllowPlaintextHttp": false
    }
  }
}

A Servers[] array may mix both transports. Each entry is wired independently; a single server’s startup failure is logged and skipped (the rest still load).

Chat

Llm (router-level guards)

Per-provider model settings do not live here (see Configuring the Per-Provider LLM Plugin Settings). These are router-level guards.

Llm:Admission (per-tenant rate-limit + concurrency; AIHost is single-tenant today)

Llm:Limits (input-payload guard)

Llm:CircuitBreaker 

Process-local breaker keyed by provider+model:

Llm:Routing

Llm:AcceptedCredentialSources

ContextCompression (proactive history compression)

Trigger: effectiveLimit = (int)(activeWindow * CompressionTriggerRatio) - ReservedResponseTokens, sized against the active model window (ILLMService.GetActiveModelCapabilities()).

How the parameters interact

Reading the knobs against the graph:

• Enabled is the top gate — false short-circuits everything.
• UseActiveModelWindow + MaxContextTokens decide which window feeds the limit.
• CompressionTriggerRatio + ReservedResponseTokens set when compression fires (lower ratio / higher reserve → fires earlier).
• MaxMessageCount is a parallel count-based trigger (fires even if tokens stay under the limit).
• CompressToolCalls decides whether tool-call turns can be summarized (false = tool exchanges always kept verbatim, never compressed).
• RecentMessageCount sets how much is preserved verbatim; TargetCompressionRatio + MaxSummaryTokens size the summary that replaces the rest.

LLMRequestSizing (input-estimate buffering)

bufferedTokens = (int)(rawTokens * (1 + InputTokenBufferRatio)) + MinBufferTokens, clamped >= rawTokens. The buffered estimate becomes LLMGenerationOptions.EstimatedInputTokens used by the router’s window gate.

What's Next? 

Refer to Configuration Manual for WeaverAI to determine if any additional configuration remains to be done.