Merge pull request #280 from drivecore/feature/local-openai-support

feature: local openai support

Author: bhouston

README.md

@@ -83,8 +83,8 @@ export default {
   profile: false,
   tokenCache: true,
 
-  // Ollama configuration (if using local models)
-  ollamaBaseUrl: 'http://localhost:11434',
+  // Base URL configuration (for providers that need it)
+  baseUrl: 'http://localhost:11434', // Example for Ollama
 };
 ```
 

mycoder.config.js

@@ -18,6 +18,9 @@ export default {
   //model: 'llama3.2:3b',
   //provider: 'xai',
   //model: 'grok-2-latest',
+  //provider: 'openai',
+  //model: 'qwen2.5-coder:14b',
+  //baseUrl: 'http://192.168.2.66:80/v1-openai',
   maxTokens: 4096,
   temperature: 0.7,
 

packages/agent/src/core/llm/__tests__/openai.test.ts

@@ -177,7 +177,7 @@ describe('OpenAIProvider', () => {
       'role' in toolUseMessage
     ) {
       expect(toolUseMessage.role).toBe('assistant');
-      expect(toolUseMessage.content).toBe(null);
+      expect(toolUseMessage.content).toBe(''); // required by gpustack' implementation of openai SDK.
 
       if (
         'tool_calls' in toolUseMessage &&

packages/agent/src/core/llm/provider.ts

@@ -57,12 +57,6 @@ export const providerConfig: Record<string, ProviderConfig> = {
     model: 'gpt-4o-2024-05-13',
     factory: (model, options) => new OpenAIProvider(model, options),
   },
-  gpustack: {
-    docsUrl: 'https://mycoder.ai/docs/provider/local-openai',
-    model: 'llama3.2',
-    baseUrl: 'http://localhost:80',
-    factory: (model, options) => new OpenAIProvider(model, options),
-  },
   ollama: {
     docsUrl: 'https://mycoder.ai/docs/provider/ollama',
     model: 'llama3.2',

packages/agent/src/core/llm/providers/openai.ts

@@ -154,7 +154,7 @@ export class OpenAIProvider implements LLMProvider {
           // so we'll include it as a function call in an assistant message
           return {
             role: 'assistant',
-            content: null,
+            content: '',
             tool_calls: [
               {
                 id: msg.id,

packages/agent/src/tools/getTools.ts

@@ -10,7 +10,6 @@ import { fetchTool } from './io/fetch.js';
 import { textEditorTool } from './io/textEditor.js';
 import { createMcpTool } from './mcp.js';
 import { listBackgroundToolsTool } from './system/listBackgroundTools.js';
-import { respawnTool } from './system/respawn.js';
 import { sequenceCompleteTool } from './system/sequenceComplete.js';
 import { shellMessageTool } from './system/shellMessage.js';
 import { shellStartTool } from './system/shellStart.js';
@@ -39,7 +38,7 @@ export function getTools(options?: GetToolsOptions): Tool[] {
     shellMessageTool as unknown as Tool,
     browseStartTool as unknown as Tool,
     browseMessageTool as unknown as Tool,
-    respawnTool as unknown as Tool,
+    //respawnTool as unknown as Tool,  this is a confusing tool for now.
     sleepTool as unknown as Tool,
     listBackgroundToolsTool as unknown as Tool,
   ];

packages/agent/src/tools/system/respawn.test.ts

@@ -8,9 +8,8 @@ import { respawnTool } from './respawn';
 const toolContext: ToolContext = getMockToolContext();
 
 describe('respawnTool', () => {
-  it('should have correct name and description', () => {
+  it('should have correct name', () => {
     expect(respawnTool.name).toBe('respawn');
-    expect(respawnTool.description).toContain('Resets the agent context');
   });
 
   it('should execute and return confirmation message', async () => {

packages/agent/src/tools/system/respawn.ts

@@ -20,7 +20,7 @@ const returnSchema = z.object({
 export const respawnTool: Tool = {
   name: 'respawn',
   description:
-    'Resets the agent context to just the system prompt and provided context',
+    'Resets the current conversation to just the system prompt and provided input context to this tool.',
   logPrefix: '🔄',
   parameters: parameterSchema,
   returns: returnSchema,

packages/cli/src/commands/$default.ts

@@ -114,18 +114,25 @@ export async function executePrompt(
       throw new Error(`Unknown provider: ${config.provider}`);
     }
 
-    const { keyName } = providerSettings;
+    // only validate key if baseUrl is not set, otherwise we assume the user is using a local provider
     let apiKey: string | undefined = undefined;
+    const { keyName } = providerSettings;
     if (keyName) {
       // Then fall back to environment variable
+      logger.info(`Looking API key in env: ${keyName}`);
       apiKey = process.env[keyName];
-      if (!apiKey) {
-        logger.error(getProviderApiKeyError(config.provider));
-        throw new Error(`${config.provider} API key not found`);
+      if (!config.baseUrl) {
+        if (!apiKey) {
+          logger.error(getProviderApiKeyError(config.provider));
+          throw new Error(`${config.provider} API key not found`);
+        }
       }
     }
 
     logger.info(`LLM: ${config.provider}/${config.model}`);
+    if (apiKey) {
+      logger.info(`Using API key: ${apiKey.slice(0, 4)}...`);
+    }
     if (config.baseUrl) {
       // For Ollama, we check if the base URL is set
       logger.info(`Using base url: ${config.baseUrl}`);

packages/docs/docs/providers/anthropic.md

@@ -32,9 +32,6 @@ export default {
   provider: 'anthropic',
   model: 'claude-3-7-sonnet-20250219',
 
-  // Optional: Set API key directly (environment variable is preferred)
-  // anthropicApiKey: 'your_api_key_here',
-
   // Other MyCoder settings
   maxTokens: 4096,
   temperature: 0.7,

packages/docs/docs/providers/index.mdx

@@ -11,8 +11,9 @@ MyCoder supports multiple Language Model (LLM) providers, giving you flexibility
 MyCoder currently supports the following LLM providers:
 
 - [**Anthropic**](./anthropic.md) - Claude models from Anthropic
-- [**OpenAI**](./openai.md) - GPT models from OpenAI
+- [**OpenAI**](./openai.md) - GPT models from OpenAI (and OpenAI compatible providers)
 - [**Ollama**](./ollama.md) - Self-hosted open-source models via Ollama
+- [**xAI**](./xai.md) - Grok models from xAI
 
 ## Configuring Providers
 
@@ -52,3 +53,4 @@ For detailed instructions on setting up each provider, see the provider-specific
 - [Anthropic Configuration](./anthropic.md)
 - [OpenAI Configuration](./openai.md)
 - [Ollama Configuration](./ollama.md)
+- [xAI Configuration](./xai.md)

packages/docs/docs/providers/ollama.md

@@ -62,7 +62,7 @@ export default {
   model: 'medragondot/Sky-T1-32B-Preview:latest',
 
   // Optional: Custom base URL (defaults to http://localhost:11434)
-  // ollamaBaseUrl: 'http://localhost:11434',
+  // baseUrl: 'http://localhost:11434',
 
   // Other MyCoder settings
   maxTokens: 4096,

packages/docs/docs/providers/openai.md

@@ -38,10 +38,6 @@ export default {
   provider: 'openai',
   model: 'gpt-4o',
 
-  // Optional: Set API key directly (environment variable is preferred)
-  // openaiApiKey: 'your_api_key_here',
-  // openaiOrganization: 'your_organization_id',
-
   // Other MyCoder settings
   maxTokens: 4096,
   temperature: 0.7,
@@ -60,6 +56,24 @@ MyCoder supports all OpenAI models that have tool/function calling capabilities.
 
 You can use any other OpenAI model that supports function calling with MyCoder. The OpenAI provider is not limited to just these listed models.
 
+## Using OpenAI Compatible Providers
+
+A number of providers offer OpenAI compatible REST API endpoints, such as xAI and [GPUStack](https://gpustack.ai). To point the OpenAI provider to a different provider REST API set the `baseUrl` and also, if applicable, the `OPENAI_API_KEY` to their required key. For example:
+
+```javascript
+export default {
+  // Provider selection
+  provider: 'openai',
+  model: 'qwen2.5',
+  baseUrl: 'http://localhost/v1-openai',
+
+  // Other MyCoder settings
+  maxTokens: 4096,
+  temperature: 0.7,
+  // ...
+};
+```
+
 ## Best Practices
 
 - GPT-4o provides the best balance of performance and cost for most MyCoder tasks

packages/docs/docs/providers/xai.md

@@ -0,0 +1,77 @@
+---
+sidebar_position: 6
+---
+
+# xAI (Grok)
+
+[xAI](https://x.ai/) is the company behind Grok, a powerful large language model designed to be helpful, harmless, and honest. Grok models offer strong reasoning capabilities and support for tool calling.
+
+## Setup
+
+To use Grok models with MyCoder, you need an xAI API key:
+
+1. Create an account at [xAI](https://x.ai/)
+2. Navigate to the API Keys section and create a new API key
+3. Set the API key as an environment variable or in your configuration file
+
+### Environment Variables
+
+You can set the xAI API key as an environment variable:
+
+```bash
+export XAI_API_KEY=your_api_key_here
+```
+
+### Configuration
+
+Configure MyCoder to use xAI's Grok in your `mycoder.config.js` file:
+
+```javascript
+export default {
+  // Provider selection
+  provider: 'xai',
+  model: 'grok-2-latest',
+
+  // Other MyCoder settings
+  maxTokens: 4096,
+  temperature: 0.7,
+  // ...
+};
+```
+
+## Supported Models
+
+xAI offers several Grok models with different capabilities:
+
+- `grok-2-latest` (recommended) - The latest Grok-2 model with strong reasoning and tool-calling capabilities
+- `grok-1` - The original Grok model
+
+## Best Practices
+
+- Grok models excel at coding tasks and technical problem-solving
+- They have strong tool-calling capabilities, making them suitable for MyCoder workflows
+- For complex programming tasks, use Grok-2 models for best results
+- Provide clear, specific instructions for optimal results
+
+## Custom Base URL
+
+If you need to use a different base URL for the xAI API (for example, if you're using a proxy or if xAI changes their API endpoint), you can specify it in your configuration:
+
+```javascript
+export default {
+  provider: 'xai',
+  model: 'grok-2-latest',
+  baseUrl: 'https://api.x.ai/v1', // Default xAI API URL
+};
+```
+
+## Troubleshooting
+
+If you encounter issues with xAI's Grok:
+
+- Verify your API key is correct and has sufficient quota
+- Check that you're using a supported model name
+- For tool-calling issues, ensure your functions are properly formatted
+- Monitor your token usage to avoid unexpected costs
+
+For more information, visit the [xAI Documentation](https://x.ai/docs).