Merge pull request #104 from leonvanzyl/ollama-support

add ollama support

.env.example

@@ -19,3 +19,20 @@
 # ANTHROPIC_DEFAULT_SONNET_MODEL=glm-4.7
 # ANTHROPIC_DEFAULT_OPUS_MODEL=glm-4.7
 # ANTHROPIC_DEFAULT_HAIKU_MODEL=glm-4.5-air
+
+# Ollama Local Model Configuration (Optional)
+# To use local models via Ollama instead of Claude, uncomment and set these variables.
+# Requires Ollama v0.14.0+ with Anthropic API compatibility.
+# See: https://ollama.com/blog/claude
+#
+# ANTHROPIC_BASE_URL=http://localhost:11434
+# ANTHROPIC_AUTH_TOKEN=ollama
+# API_TIMEOUT_MS=3000000
+# ANTHROPIC_DEFAULT_SONNET_MODEL=qwen3-coder
+# ANTHROPIC_DEFAULT_OPUS_MODEL=qwen3-coder
+# ANTHROPIC_DEFAULT_HAIKU_MODEL=qwen3-coder
+#
+# Model recommendations:
+# - For best results, use a capable coding model like qwen3-coder or deepseek-coder-v2
+# - You can use the same model for all tiers, or different models per tier
+# - Larger models (70B+) work best for Opus tier, smaller (7B-20B) for Haiku

.gitignore

@@ -76,6 +76,11 @@ ui/playwright-report/
 .dmypy.json
 dmypy.json
 
+# ===================
+# Claude Code
+# ===================
+.claude/settings.local.json
+
 # ===================
 # IDE / Editors
 # ===================

CLAUDE.md

@@ -256,6 +256,39 @@ python test_security_integration.py
 - `examples/README.md` - Comprehensive guide with use cases, testing, and troubleshooting
 - `PHASE3_SPEC.md` - Specification for mid-session approval feature (future enhancement)
 
+### Ollama Local Models (Optional)
+
+Run coding agents using local models via Ollama v0.14.0+:
+
+1. Install Ollama: https://ollama.com
+2. Start Ollama: `ollama serve`
+3. Pull a coding model: `ollama pull qwen3-coder`
+4. Configure `.env`:
+   ```
+   ANTHROPIC_BASE_URL=http://localhost:11434
+   ANTHROPIC_AUTH_TOKEN=ollama
+   API_TIMEOUT_MS=3000000
+   ANTHROPIC_DEFAULT_SONNET_MODEL=qwen3-coder
+   ANTHROPIC_DEFAULT_OPUS_MODEL=qwen3-coder
+   ANTHROPIC_DEFAULT_HAIKU_MODEL=qwen3-coder
+   ```
+5. Run autocoder normally - it will use your local Ollama models
+
+**Recommended coding models:**
+- `qwen3-coder` - Good balance of speed and capability
+- `deepseek-coder-v2` - Strong coding performance
+- `codellama` - Meta's code-focused model
+
+**Model tier mapping:**
+- Use the same model for all tiers, or map different models per capability level
+- Larger models (70B+) work best for Opus tier
+- Smaller models (7B-20B) work well for Haiku tier
+
+**Known limitations:**
+- Smaller context windows than Claude (model-dependent)
+- Extended context beta disabled (not supported by Ollama)
+- Performance depends on local hardware (GPU recommended)
+
 ## Claude Code Integration
 
 - `.claude/commands/create-spec.md` - `/create-spec` slash command for interactive spec creation

client.py

@@ -257,9 +257,16 @@ def create_client(
         if value:
             sdk_env[var] = value
 
+    # Detect alternative API mode (Ollama or GLM)
+    base_url = sdk_env.get("ANTHROPIC_BASE_URL", "")
+    is_alternative_api = bool(base_url)
+    is_ollama = "localhost:11434" in base_url or "127.0.0.1:11434" in base_url
+
     if sdk_env:
         print(f"   - API overrides: {', '.join(sdk_env.keys())}")
-        if "ANTHROPIC_BASE_URL" in sdk_env:
+        if is_ollama:
+            print("   - Ollama Mode: Using local models")
+        elif "ANTHROPIC_BASE_URL" in sdk_env:
             print(f"   - GLM Mode: Using {sdk_env['ANTHROPIC_BASE_URL']}")
 
     # Create a wrapper for bash_security_hook that passes project_dir via context
@@ -336,7 +343,8 @@ def create_client(
             # Enable extended context beta for better handling of long sessions.
             # This provides up to 1M tokens of context with automatic compaction.
             # See: https://docs.anthropic.com/en/api/beta-headers
-            betas=["context-1m-2025-08-07"],
+            # Disabled for alternative APIs (Ollama, GLM) as they don't support Claude-specific betas.
+            betas=[] if is_alternative_api else ["context-1m-2025-08-07"],
             # Note on context management:
             # The Claude Agent SDK handles context management automatically through the
             # underlying Claude Code CLI. When context approaches limits, the CLI

server/routers/settings.py

@@ -40,7 +40,15 @@ def _parse_yolo_mode(value: str | None) -> bool:
 
 def _is_glm_mode() -> bool:
     """Check if GLM API is configured via environment variables."""
-    return bool(os.getenv("ANTHROPIC_BASE_URL"))
+    base_url = os.getenv("ANTHROPIC_BASE_URL", "")
+    # GLM mode is when ANTHROPIC_BASE_URL is set but NOT pointing to Ollama
+    return bool(base_url) and not _is_ollama_mode()
+
+
+def _is_ollama_mode() -> bool:
+    """Check if Ollama API is configured via environment variables."""
+    base_url = os.getenv("ANTHROPIC_BASE_URL", "")
+    return "localhost:11434" in base_url or "127.0.0.1:11434" in base_url
 
 
 @router.get("/models", response_model=ModelsResponse)
@@ -82,6 +90,7 @@ async def get_settings():
         yolo_mode=_parse_yolo_mode(all_settings.get("yolo_mode")),
         model=all_settings.get("model", DEFAULT_MODEL),
         glm_mode=_is_glm_mode(),
+        ollama_mode=_is_ollama_mode(),
         testing_agent_ratio=_parse_int(all_settings.get("testing_agent_ratio"), 1),
     )
 
@@ -104,5 +113,6 @@ async def update_settings(update: SettingsUpdate):
         yolo_mode=_parse_yolo_mode(all_settings.get("yolo_mode")),
         model=all_settings.get("model", DEFAULT_MODEL),
         glm_mode=_is_glm_mode(),
+        ollama_mode=_is_ollama_mode(),
         testing_agent_ratio=_parse_int(all_settings.get("testing_agent_ratio"), 1),
     )

server/schemas.py

@@ -382,6 +382,7 @@ class SettingsResponse(BaseModel):
     yolo_mode: bool = False
     model: str = DEFAULT_MODEL
     glm_mode: bool = False  # True if GLM API is configured via .env
+    ollama_mode: bool = False  # True if Ollama API is configured via .env
     testing_agent_ratio: int = 1  # Regression testing agents (0-3)
 
 

ui/src/App.tsx

@@ -298,6 +298,17 @@ function App() {
                     <Settings size={18} />
                   </button>
 
+                  {/* Ollama Mode Indicator */}
+                  {settings?.ollama_mode && (
+                    <div
+                      className="flex items-center gap-1.5 px-2 py-1 bg-white rounded border-2 border-neo-border shadow-neo-sm"
+                      title="Using Ollama local models (configured via .env)"
+                    >
+                      <img src="/ollama.png" alt="Ollama" className="w-5 h-5" />
+                      <span className="text-xs font-bold text-neo-text">Ollama</span>
+                    </div>
+                  )}
+
                   {/* GLM Mode Badge */}
                   {settings?.glm_mode && (
                     <span

ui/src/hooks/useProjects.ts

@@ -237,6 +237,7 @@ const DEFAULT_SETTINGS: Settings = {
   yolo_mode: false,
   model: 'claude-opus-4-5-20251101',
   glm_mode: false,
+  ollama_mode: false,
   testing_agent_ratio: 1,
 }
 

ui/src/lib/types.ts

@@ -526,6 +526,7 @@ export interface Settings {
   yolo_mode: boolean
   model: string
   glm_mode: boolean
+  ollama_mode: boolean
   testing_agent_ratio: number  // Regression testing agents (0-3)
 }