diff --git a/README.md b/README.md
index 91bdf20..a6c5eb1 100644
--- a/README.md
+++ b/README.md
@@ -125,7 +125,35 @@ You're done!
   - I've pre-loaded Pyrite on the server. Just set your url to `http://localhost:42069/v1/pyrite`! This is meant for people who JUST installed a front and and don't have a real setup yet - it's nice to be able to celebrate your victory with something working right away!
 - Read up on how SillyTavern handles caching: https://docs.sillytavern.app/administration/config-yaml/#claude-configuration
   - It's off by default, turn it on with those configs. Choose depth 0 if you aren't sure; this caches the most aggressively.
-  - What all those warnings mean is that for cache to be used, the convo history up to a certain point has to be the exact same. ST has a lot of advanced features where it makes changes to the start of the context, ruining your savings. But for simpler use cases, it's fine. Set the context to 200K IMO - as stuff falls out of context if you choose a lower number, that changes the convo start 
+  - What all those warnings mean is that for cache to be used, the convo history up to a certain point has to be the exact same. ST has a lot of advanced features where it makes changes to the start of the context, ruining your savings. But for simpler use cases, it's fine. Set the context to 200K IMO - as stuff falls out of context if you choose a lower number, that changes the convo start
+
+### TypingMind Setup
+
+Unlike SillyTavern, TypingMind requires the **full endpoint URL** including `/v1/messages`.
+
+1. In TypingMind, go to **Model** > **Add Custom Model**
+2. Set **AI Type** to `Claude (messages API)`
+3. Set **Endpoint URL** to `http://localhost:42069/v1/messages` (note: full path required)
+4. Set **Model ID** to a valid Claude model, e.g. `claude-opus-4-5-20250514` or `claude-sonnet-4-20250514`
+5. Set **Context Length** to `200000` (same for all Claude models)
+6. Leave **Authentication** set to `None`
+7. (Optional) Add a PHI instruction - see below
+8. Click **Test**, then **Add Model** when the test succeeds
+
+#### Adding PHI (Post History Instruction)
+
+PHI allows you to inject ephemeral instructions that appear after the last user message but aren't stored in conversation history. This is useful for reply rules, formatting preferences, or per-request instructions.
+
+1. Expand **Advanced** > **Custom Body Params**
+2. Click **Add Body Parameter**
+3. Set type to `string`, key to `phi`, and value to your instruction
+
+Example PHI value:
+```
+<replyRules>Always respond concisely. Use markdown formatting.</replyRules>
+```
+
+The proxy strips the `phi` field from the request and injects it as a user message after your last message. Claude sees it, but TypingMind doesn't store it in the conversation history.
 
 ### Troubleshooting
 
@@ -155,6 +183,7 @@ Most likely thing to go wrong is not being able to find the credentials, either
 - Remove "ttl" key from any "cache_control" objects, since endpoint does not allow it
 - The first section of the system prompt must be "You are Claude Code, Anthropic's official CLI for Claude." or the request will not be accepted by Anthropic (specifically/technically, it must be the first item of the "system" array's "text" content). I am adding this, but this is just FYI so you know it's there and that you have to deal with it
 - Optionally filter sampling parameters to avoid conflicts with Sonnet 4.5. Set `filter_sampling_params=true` in `server/config.txt` to enable this feature, which ensures only one sampling parameter is sent to the API. When both `temperature` and `top_p` are specified, it removes whichever is at the default value (1.0), or prefers temperature if both are non-default (Sonnet 4.5 doesn't allow both parameters). Other models work fine with both parameters, so this defaults to off
+- **PHI (Post History Instruction)**: If the request body contains a `phi` field, it's injected as a user message after the last user message, then removed from the request. This allows clients like TypingMind to send ephemeral instructions without native PHI support
 
 ### Smart Host Binding
 - **Native execution**: Binds to `127.0.0.1` (secure, local-only)
diff --git a/server/ClaudeRequest.js b/server/ClaudeRequest.js
index 4a468d8..c8905fd 100644
--- a/server/ClaudeRequest.js
+++ b/server/ClaudeRequest.js
@@ -48,7 +48,7 @@ class ClaudeRequest {
   constructor(req = null) {
     this.API_URL = 'https://api.anthropic.com/v1/messages';
     this.VERSION = '2023-06-01';
-    this.BETA_HEADER = 'claude-code-20250219,oauth-2025-04-20,interleaved-thinking-2025-05-14,fine-grained-tool-streaming-2025-05-14';
+    this.BETA_HEADER = 'claude-code-20250219,oauth-2025-04-20,interleaved-thinking-2025-05-14,fine-grained-tool-streaming-2025-05-14,web-search-2025-03-05,web-fetch-2025-09-10';
 
     const apiKey = req?.headers?.['x-api-key'];
     if (apiKey && apiKey.includes('sk-ant')) {
@@ -225,7 +225,7 @@ class ClaudeRequest {
     if (ClaudeRequest.refreshPromise) {
       return await ClaudeRequest.refreshPromise;
     }
-    
+
     ClaudeRequest.refreshPromise = this._doRefresh();
     try {
       const result = await ClaudeRequest.refreshPromise;
@@ -286,20 +286,20 @@ class ClaudeRequest {
         req.write(JSON.stringify(refreshData));
         req.end();
       });
-      
+
       credentials.claudeAiOauth.accessToken = response.access_token;
       credentials.claudeAiOauth.refreshToken = response.refresh_token;
       credentials.claudeAiOauth.expiresAt = Date.now() + (response.expires_in * 1000);
-      
+
       const credentialsJson = JSON.stringify(credentials);
       this.writeCredentialsToFile(credentialsJson);
-      
+
       Logger.info('Token refreshed successfully');
       return `Bearer ${response.access_token}`;
-      
+
     } catch (error) {
       if (error.code === 'ENOENT') {
-        const errorMsg = process.platform === 'win32' 
+        const errorMsg = process.platform === 'win32'
           ? 'Failed to load credentials: Claude credentials file not found in WSL. Check your default WSL distro with "wsl -l -v" and set the correct one with "wsl --set-default <distro-name>". As a backup, you can get the token from ~/.claude/.credentials.json and pass it as x-api-key (proxy password in SillyTavern)'
           : 'Claude credentials not found. Please ensure Claude Code is installed and you have logged in. As a backup, you can get the token from ~/.claude/.credentials.json and pass it as x-api-key (proxy password in SillyTavern)';
         Logger.error('ENOENT error during token refresh:', errorMsg);
@@ -342,7 +342,7 @@ class ClaudeRequest {
       if (Array.isArray(body.system)) {
         body.system.unshift(systemPrompt);
       } else {
-        body.system = [systemPrompt, body.system];
+        body.system = [systemPrompt, { type: 'text', text: body.system }];
       }
     } else {
       body.system = [systemPrompt];
@@ -352,6 +352,10 @@ class ClaudeRequest {
       this.applyPreset(body, presetName);
     }
 
+    if (body.phi !== undefined || body.PHI !== undefined) {
+      this.injectPHI(body);
+    }
+
     body = this.stripTtlFromCacheControl(body);
     body = this.filterSamplingParams(body);
 
@@ -394,7 +398,7 @@ class ClaudeRequest {
     // Use suffixEt only when thinking is enabled, otherwise use regular suffix
     const hasThinking = body.thinking && body.thinking.type === 'enabled';
     const suffix = hasThinking ? preset.suffixEt : preset.suffix;
-    
+
     if (suffix && body.messages && body.messages.length > 0) {
       const lastUserIndex = body.messages.map(m => m.role).lastIndexOf('user');
       if (lastUserIndex !== -1) {
@@ -409,6 +413,24 @@ class ClaudeRequest {
     Logger.debug(`Applied preset: ${presetName}`);
   }
 
+  injectPHI(body) {
+    const phi = body.phi || body.PHI;
+    delete body.phi;  // Always remove before sending to Claude API
+    delete body.PHI;
+
+    if (phi && body.messages && body.messages.length > 0) {
+      const lastUserIndex = body.messages.map(m => m.role).lastIndexOf('user');
+      if (lastUserIndex !== -1) {
+        const phiMsg = {
+          role: 'user',
+          content: [{ type: 'text', text: phi }]
+        };
+        body.messages.splice(lastUserIndex + 1, 0, phiMsg);
+        Logger.debug('Injected PHI message');
+      }
+    }
+  }
+
   async makeRequest(body, presetName = null) {
     const token = await this.getAuthToken();
     const headers = this.getHeaders(token);
@@ -435,7 +457,7 @@ class ClaudeRequest {
         req.destroy();
         reject(err);
       });
-      
+
       req.write(JSON.stringify(processedBody));
       req.end();
     });
@@ -444,11 +466,11 @@ class ClaudeRequest {
   async handleResponse(res, body, presetName = null) {
     try {
       const claudeResponse = await this.makeRequest(body, presetName);
-      
+
       if (claudeResponse.statusCode === 401) {
         Logger.info('Got 401, checking credential store');
         ClaudeRequest.cachedToken = null;
-        
+
         try {
           const newToken = await this.loadOrRefreshToken();
           ClaudeRequest.cachedToken = newToken;
@@ -465,16 +487,16 @@ class ClaudeRequest {
           Logger.info('Token load/refresh failed, passing 401 to client');
         }
       }
-      
+
       res.statusCode = claudeResponse.statusCode;
       Logger.debug(`Claude API status: ${claudeResponse.statusCode}`);
       Logger.debug('Claude response headers:', JSON.stringify(claudeResponse.headers, null, 2));
       Object.keys(claudeResponse.headers).forEach(key => {
         res.setHeader(key, claudeResponse.headers[key]);
       });
-      
+
       this.streamResponse(res, claudeResponse);
-      
+
     } catch (error) {
       console.error('Claude request error:', error.message);
       res.writeHead(500, { 'Content-Type': 'application/json' });
@@ -507,7 +529,7 @@ class ClaudeRequest {
     const contentType = claudeResponse.headers['content-type'] || '';
     if (contentType.includes('text/event-stream')) {
       Logger.debug('Outgoing response headers to client:', JSON.stringify(res.getHeaders(), null, 2));
-      
+
       claudeResponse.on('error', (err) => {
         Logger.debug('Claude response stream error:', err);
         if (!res.headersSent) {
@@ -517,17 +539,17 @@ class ClaudeRequest {
           res.end(JSON.stringify({ error: 'Upstream response error' }));
         }
       });
-      
+
       res.on('close', () => {
         Logger.debug('Client disconnected, cleaning up streams');
         if (!claudeResponse.destroyed) {
           claudeResponse.destroy();
         }
       });
-      
+
       if (Logger.getLogLevel() >= 3) {
         const debugStream = Logger.createDebugStream('Claude SSE', extractClaudeText);
-        
+
         debugStream.on('error', (err) => {
           Logger.debug('Debug stream error:', err);
           if (!res.headersSent) {
@@ -537,7 +559,7 @@ class ClaudeRequest {
             res.end(JSON.stringify({ error: 'Stream processing error' }));
           }
         });
-        
+
         claudeResponse.pipe(debugStream).pipe(res);
         debugStream.on('end', () => {
           Logger.debug('\n');
diff --git a/server/server.js b/server/server.js
index b4c5bc8..e3ac92c 100644
--- a/server/server.js
+++ b/server/server.js
@@ -127,7 +127,7 @@ async function handleRequest(req, res) {
 
   res.setHeader('Access-Control-Allow-Origin', '*');
   res.setHeader('Access-Control-Allow-Methods', 'GET, POST, PUT, DELETE, OPTIONS');
-  res.setHeader('Access-Control-Allow-Headers', 'Content-Type, Authorization, X-Requested-With');
+  res.setHeader('Access-Control-Allow-Headers', 'Content-Type, Authorization, X-Requested-With, anthropic-version, anthropic-beta, x-api-key');
 
   if (req.method === 'OPTIONS') {
     res.writeHead(200);