Ollama Turbo (Cloud) Compatibility

[gbsierra]

Ollama Turbo Compatibility Fix Plan

Issue

Users cannot use Ollama Turbo (cloud service) with BrowserOS because:

• ollama type forces localhost and lacks cloud API key support
• openai_compatible and custom types force /v1 path and use wrong client
• No existing provider handles Ollama Turbo's requirements properly

Current State

Provider Type	Client	Base URL	API Key	Works for Ollama Turbo?
ollama	ChatOllama	localhost:11434	❌ No	❌ (localhost only)
openai_compatible	ChatOpenAI	forces /v1	✅ Yes	❌ (wrong client/format)
custom	ChatOpenAI	forces /v1	✅ Yes	❌ (same as above)

Proposed Solution

Nothing changes for existing users. Only when users select ollama type and set a non-localhost base URL, then API key support is enabled.

Detection Logic

private _isLocalOllama(baseUrl: string): boolean {
  if (!baseUrl) return true // Default is local (localhost:11434)

  return baseUrl.includes('localhost') ||
         baseUrl.includes('127.0.0.1') ||
         baseUrl.includes('0.0.0.0') ||
         !baseUrl.includes('://') // relative URLs like 'localhost:8080'
}

// Usage:
if (this._isLocalOllama(provider.baseUrl)) {
  // Local Ollama - current behavior (no API key needed)
} else {
  // Cloud Ollama - add API key support
}

Detection Examples:

Base URL	Detected As	Behavior
"" (blank)	Local ✅	Uses localhost:11434, no API key
"localhost:11434"	Local ✅	No API key
"127.0.0.1:8080"	Local ✅	No API key
"https://ollama.com"	Cloud 🚀	Enables API key support
"https://api.ollama.com"	Cloud 🚀	Enables API key support
"192.168.1.100:11434"	Cloud 🚀	Enables API key support

Implementation

• Existing local users: Zero changes to workflow
• New cloud users: Set non-localhost URL + API key
• Smart detection: Automatic based on URL
• Backward compatibility: 100%

User Experience

Local Ollama (unchanged):

Type: ollama
Base URL: http://localhost:11434 (or blank)
API Key: (not required)

Ollama Turbo (new capability):

Type: ollama
Base URL: https://ollama.com
API Key: [user's Ollama Turbo key]

Benefits

• ✅ No new provider types needed
• ✅ Backward compatible with existing local Ollama users
• ✅ Clear, intuitive user experience
• ✅ Extensible for other cloud Ollama services
• ✅ Fixes Ollama Turbo compatibility

Implementation Steps

1. Add localhost detection logic to _createOllamaLLM
2. Add API key header support for cloud services
3. Test with both local and cloud Ollama endpoints
4. Update documentation for users

Risks

• Edge cases with local network Ollama instances
• API format differences between local and cloud Ollama
• Need to verify Ollama Turbo's exact auth requirements

[gbsierra]

Updated Ollama Turbo Implementation Status

✅ What We Know Works

• API connection to https://ollama.com
• Authentication with Bearer token

🚨 Issues Identified & Solutions Planned

Issue 1: Provider Creation (Base URL & API Key)

• Problem: Need conditional logic to detect Ollama Turbo vs local Ollama
• Solution: Modify _createOllamaLLM in LangChainProvider.ts to detect cloud vs local

Issue 2: Tool Name Template Token Leakage

• Problem: Ollama Turbo leaks template tokens like <|channel|>analysis into tool names
• Possible Solution: Implement tool name sanitization in BrowserAgent.ts - ONLY for Ollama Turbo

Issue 3: Message History Validation Errors

• Problem: Ollama Turbo has stricter validation than local Ollama
• Possible Solution: Implement message history sanitization in BrowserAgent.ts - ONLY for Ollama Turbo

Issue 4: ResultTool Output Format Differences

• Problem: Ollama Turbo returns different response format than other providers
• Possible Solution: Implement conditional parsing in ResultTool.ts - ONLY for Ollama Turbo

CONDITIONAL IMPLEMENTATION PLAN - SHOULD HAVE ZERO IMPACT ON EXISTING CODE

Phase 1: Provider Detection & Creation

File: src/lib/llm/LangChainProvider.ts
Goal: Detect Ollama Turbo vs local Ollama, configure appropriately

Phase 2: Tool Name Sanitization - CONDITIONAL ONLY

File: src/lib/agent/BrowserAgent.ts
Goal: Fix the "Invalid tool usage: mismatch between tool calls and tool results" error

Phase 3: Message History Sanitization - CONDITIONAL ONLY

File: src/lib/agent/BrowserAgent.ts
Goal: Prevent Ollama Turbo validation errors from problematic messages

Phase 4: ResultTool Conditional Parsing - CONDITIONAL ONLY

File: src/lib/tools/result/ResultTool.ts
Goal: Handle Ollama Turbo's different output format WITHOUT changing other providers

Implementation Strategy

Implementation Order

1. Phase 1: Provider detection only
2. Phase 2: Tool name sanitization
3. Phase 3: Message history sanitization
4. Phase 4: ResultTool conditional (minimal changes only)

Expected Outcomes

• All existing providers work exactly as before - NO REGRESSIONS
• Ollama Turbo works conditionally

📚 Reference

• Ollama Turbo GitHub Issue #12132 - bug report confirming token leakage issue