diff --git a/mcp/src/aieo/src/provider.ts b/mcp/src/aieo/src/provider.ts
index fa9d1d53d..b54ccb118 100644
--- a/mcp/src/aieo/src/provider.ts
+++ b/mcp/src/aieo/src/provider.ts
@@ -356,9 +356,23 @@ export function getTokenPricing(provider: Provider): TokenPricing {
 
 export type ThinkingSpeed = "thinking" | "fast";
 
+// Anthropic models that do NOT support `adaptive` thinking. Adaptive is only
+// available on certain Sonnet/Opus tiers; Haiku and older models reject it
+// with `invalid_request_error: adaptive thinking is not supported on this model`.
+function anthropicSupportsAdaptiveThinking(modelName?: string): boolean {
+  if (!modelName) return false;
+  const m = modelName.toLowerCase();
+  // Haiku family does not support adaptive thinking.
+  if (m.includes("haiku")) return false;
+  // Conservatively only enable adaptive thinking for known-supported families.
+  // Sonnet 4.5+ and Opus 4.5+ support it.
+  return m.includes("sonnet") || m.includes("opus");
+}
+
 export function getProviderOptions(
   provider: Provider,
-  thinkingSpeed?: ThinkingSpeed
+  thinkingSpeed?: ThinkingSpeed,
+  modelName?: string
 ) {
   const fast = thinkingSpeed === "fast";
   const explicitThinking = thinkingSpeed === "thinking";
@@ -372,9 +386,12 @@ export function getProviderOptions(
         thinking = { type: "disabled" };
       } else if (explicitThinking) {
         thinking = { type: "enabled", budgetTokens: 24000 };
-      } else {
-        // Default: let the model decide, with summarized thinking output.
+      } else if (anthropicSupportsAdaptiveThinking(modelName)) {
+        // Default for capable models: let the model decide, with summarized thinking output.
         thinking = { type: "adaptive", display: "summarized" };
+      } else {
+        // Models that don't support adaptive thinking (e.g. Haiku): disable it.
+        thinking = { type: "disabled" };
       }
       return {
         anthropic: {
diff --git a/mcp/src/aieo/src/stream.ts b/mcp/src/aieo/src/stream.ts
index 6ead0b04c..dabc450c6 100644
--- a/mcp/src/aieo/src/stream.ts
+++ b/mcp/src/aieo/src/stream.ts
@@ -47,7 +47,7 @@ export async function callModel(opts: CallModelOptions): Promise<{
     executablePath,
     modelName,
   });
-  const providerOptions = getProviderOptions(provider, thinkingSpeed);
+  const providerOptions = getProviderOptions(provider, thinkingSpeed, modelName);
   console.log(`Calling ${provider} with options:`, providerOptions);
   const result = streamText({
     model,
diff --git a/mcp/src/gitsee/agent/explore.ts b/mcp/src/gitsee/agent/explore.ts
index 20aa95733..230e3da7c 100644
--- a/mcp/src/gitsee/agent/explore.ts
+++ b/mcp/src/gitsee/agent/explore.ts
@@ -157,7 +157,7 @@ export async function gitsee_context(
     tools,
     prompt,
     system: overrides?.system_prompt || CONF.system,
-    providerOptions: getProviderOptions(llm.provider) as any,
+    providerOptions: getProviderOptions(llm.provider, undefined, llm.modelName) as any,
     stopWhen: hasToolCall("final_answer"),
     onStepFinish: (sf) => logStep(sf.content),
   });
diff --git a/mcp/src/graph/learnings.ts b/mcp/src/graph/learnings.ts
index aec9eaac3..144741086 100644
--- a/mcp/src/graph/learnings.ts
+++ b/mcp/src/graph/learnings.ts
@@ -91,7 +91,7 @@ export async function post_relevant_learnings(req: Request, res: Response) {
   try {
     const llm = resolveLLMConfig({ model: req.body.model, apiKey: req.body.apiKey, light: true });
     const model = llm.model;
-    const providerOptions = getProviderOptions(llm.provider);
+    const providerOptions = getProviderOptions(llm.provider, undefined, llm.modelName);
 
     // 1. List all scopes
     const allScopes = await db.get_all_scopes();
diff --git a/mcp/src/log/agent.ts b/mcp/src/log/agent.ts
index dc5cc31dc..489137f88 100644
--- a/mcp/src/log/agent.ts
+++ b/mcp/src/log/agent.ts
@@ -59,7 +59,7 @@ export async function log_agent_context(
   opts: LogAgentOptions
 ): Promise<ContextResult> {
   const startTime = Date.now();
-  const { model, provider } = getModelDetails(opts.modelName, opts.apiKey);
+  const { model, provider, modelId } = getModelDetails(opts.modelName, opts.apiKey);
   console.log("===> log_agent model", model);
 
   const tools = get_log_tools({
@@ -83,7 +83,7 @@ export async function log_agent_context(
     model,
     instructions: SYSTEM,
     tools,
-    providerOptions: getProviderOptions(provider) as any,
+    providerOptions: getProviderOptions(provider, undefined, modelId) as any,
     stopWhen: hasEndMarker,
     stopSequences: ["[END_OF_ANSWER]"],
     onStepFinish: (sf) => {
diff --git a/mcp/src/repo/agent.ts b/mcp/src/repo/agent.ts
index ce3887cf2..cb9d858fb 100644
--- a/mcp/src/repo/agent.ts
+++ b/mcp/src/repo/agent.ts
@@ -446,8 +446,8 @@ Apply the guidance from each skill throughout your response.`;
 
 /** Build the generate/stream call params from the prepared agent state. */
 function buildCallParams(prepared: PreparedAgent) {
-  const { finalPrompt, previousMessages, userMessage, provider, abortSignal } = prepared;
-  const providerOptions = getProviderOptions(provider as any);
+  const { finalPrompt, previousMessages, userMessage, provider, modelId, abortSignal } = prepared;
+  const providerOptions = getProviderOptions(provider as any, undefined, modelId);
   const base = abortSignal ? { providerOptions, abortSignal } : { providerOptions };
   if (previousMessages.length > 0) {
     const messagesToSend =
diff --git a/mcp/src/repo/descriptions.ts b/mcp/src/repo/descriptions.ts
index f08caf1c1..65fd2876b 100644
--- a/mcp/src/repo/descriptions.ts
+++ b/mcp/src/repo/descriptions.ts
@@ -119,7 +119,7 @@ export const describe_nodes_agent = async (req: Request, res: Response) => {
     let totalProcessed = 0;
     let totalUsage = emptyUsage();
     const model = llm.model;
-    const providerOptions = getProviderOptions(llm.provider);
+    const providerOptions = getProviderOptions(llm.provider, undefined, llm.modelName);
 
     // Loop until cost limit reached or no more nodes
     while (true) {
diff --git a/mcp/src/repo/docs.ts b/mcp/src/repo/docs.ts
index 3e7c445a4..a753508a6 100644
--- a/mcp/src/repo/docs.ts
+++ b/mcp/src/repo/docs.ts
@@ -12,7 +12,7 @@ export async function learn_docs_agent(req: Request, res: Response) {
   const reqApiKey = (req.query.apiKey || req.body?.apiKey) as string | undefined;
   const llm = resolveLLMConfig({ model: reqModel, apiKey: reqApiKey });
   const model = llm.model;
-  const providerOptions = getProviderOptions(llm.provider);
+  const providerOptions = getProviderOptions(llm.provider, undefined, llm.modelName);
 
   try {
     const allRepos = await db.get_repositories();
diff --git a/mcp/src/repo/workflows.ts b/mcp/src/repo/workflows.ts
index 75692beb1..a4d2c5632 100644
--- a/mcp/src/repo/workflows.ts
+++ b/mcp/src/repo/workflows.ts
@@ -30,7 +30,7 @@ export async function document_workflow(req: Request, res: Response) {
   res.json({ request_id, status: 'pending' });
 
   const llm = resolveLLMConfig({ model: req.body.model, apiKey: req.body.apiKey });
-  const providerOptions = getProviderOptions(llm.provider);
+  const providerOptions = getProviderOptions(llm.provider, undefined, llm.modelName);
   (async () => {
     try {
       const result = await generateText({ model: llm.model, prompt: buildPrompt(workflow.workflow_json || JSON.stringify(workflow)), providerOptions: providerOptions as any });
@@ -52,7 +52,7 @@ export async function document_workflows(req: Request, res: Response) {
 
   const llm = resolveLLMConfig({ model: req.body.model, apiKey: req.body.apiKey });
   const model = llm.model;
-  const providerOptions = getProviderOptions(llm.provider);
+  const providerOptions = getProviderOptions(llm.provider, undefined, llm.modelName);
   (async () => {
     try {
       const workflows = await db.get_all_workflows();
diff --git a/mcp/src/tools/explore/tool.ts b/mcp/src/tools/explore/tool.ts
index 18cea1115..240230f46 100644
--- a/mcp/src/tools/explore/tool.ts
+++ b/mcp/src/tools/explore/tool.ts
@@ -174,7 +174,7 @@ export async function get_context_explore(
     tools,
     prompt,
     system,
-    providerOptions: getProviderOptions(llm.provider) as any,
+    providerOptions: getProviderOptions(llm.provider, undefined, llm.modelName) as any,
     stopWhen: hasToolCall("final_answer"),
     onStepFinish: (sf) => {
       // console.log("step", JSON.stringify(sf.content, null, 2));