From 65cca005c362fb14a3dfb0cd889a7e79fd8b65cb Mon Sep 17 00:00:00 2001
From: niko4244 <112509534+niko4244@users.noreply.github.com>
Date: Fri, 17 Jul 2026 13:29:01 -0500
Subject: [PATCH] feat: add local Ollama provider support with prefix stripping
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Two changes to enable using a local Ollama instance (127.0.0.1:11434)
instead of the Ollama Cloud API:

1. providers/index.ts: Change Ollama baseUrl from https://ollama.com/v1
   to http://127.0.0.1:11434/v1, and rename from 'Ollama Cloud' to
   'Local Ollama'.

2. providers/openai-compat.ts: Strip the 'ollama/' prefix from model
   IDs before sending to Ollama's API. FreeLLMAPI stores model IDs as
   'ollama/<name>' but the local Ollama API expects bare model names
   (e.g., 'qwen3.5:4b' not 'ollama/qwen3.5:4b').

Both changes are backwards-compatible — the prefix stripping is a
no-op when no prefix exists, and the base URL can be overridden via
provider config.
---
 server/src/providers/index.ts         | 23 +++++++----------------
 server/src/providers/openai-compat.ts |  8 ++++++++
 2 files changed, 15 insertions(+), 16 deletions(-)

diff --git a/server/src/providers/index.ts b/server/src/providers/index.ts
index a149c9d9c..59922751c 100644
--- a/server/src/providers/index.ts
+++ b/server/src/providers/index.ts
@@ -94,23 +94,14 @@ register(new OpenAICompatProvider({
 }));
 
 // Moonshot direct integration was dropped in V4 (paid-only); MiniMax direct
-// was dropped in V4 (superseded by the OpenRouter route).
-
-// Ollama Cloud — OpenAI-compatible. Free plan: 1 concurrent model, 5h session
-// caps, GPU-time-based quota (not per-token). Many catalog models on the
-// /v1/models list are subscription-only — Free returns 403 with an explicit
-// "this model requires a subscription" message. Catalog rows are filtered to
-// confirmed-Free entries.
-//
-// Frontier reasoning models (glm-4.7, kimi-k2-thinking, cogito-2.1:671b)
-// regularly take 30-90s on Ollama Cloud Free, so the timeout is bumped from
-// the default 15s. Ollama returns reasoning in `message.reasoning` (not
-// `reasoning_content`) — handled by normalizeChoices.
+// was dropped in V4 (superseded by the OpenRouter route).// Local Ollama — OpenAI-compatible. Runs at 127.0.0.1:11434. Models
+// are the local models pulled via `ollama pull`. No auth required.
+// Timeout matches the custom-provider extended timeout.
 register(new OpenAICompatProvider({
-  platform: 'ollama',
-  name: 'Ollama Cloud',
-  baseUrl: 'https://ollama.com/v1',
-  timeoutMs: 120000,
+    platform: 'ollama',
+    name: 'Local Ollama',
+    baseUrl: 'http://127.0.0.1:11434/v1',
+    timeoutMs: 120000,
 }));
 
 // Kilo AI Gateway — OpenAI-compatible aggregator. Anonymous access works
diff --git a/server/src/providers/openai-compat.ts b/server/src/providers/openai-compat.ts
index 23fa43a3c..e93be8215 100644
--- a/server/src/providers/openai-compat.ts
+++ b/server/src/providers/openai-compat.ts
@@ -44,6 +44,10 @@ export class OpenAICompatProvider extends BaseProvider {
     modelId: string,
     options?: CompletionOptions,
   ): Promise<ChatCompletionResponse> {
+    // Strip platform prefix for providers that use bare model names (local Ollama).
+    if (this.platform === 'ollama') {
+      modelId = modelId.replace(/^ollama\//, '');
+    }
     const res = await this.fetchWithTimeout(`${this.baseUrl}/chat/completions`, {
       method: 'POST',
       headers: {
@@ -80,6 +84,10 @@ export class OpenAICompatProvider extends BaseProvider {
     modelId: string,
     options?: CompletionOptions,
   ): AsyncGenerator<ChatCompletionChunk> {
+    // Strip platform prefix for providers that use bare model names (local Ollama).
+    if (this.platform === 'ollama') {
+      modelId = modelId.replace(/^ollama\//, '');
+    }
     const res = await this.fetchWithTimeout(`${this.baseUrl}/chat/completions`, {
       method: 'POST',
       headers: {