diff --git a/README.md b/README.md
index cdc2c0f2..e26f2232 100644
--- a/README.md
+++ b/README.md
@@ -125,7 +125,16 @@ Dimension is detected automatically. ⚠️ Switching models changes the vector
 
 ## MCP Server
 
-The `mcp-server/` directory contains a standalone [MCP](https://modelcontextprotocol.io) server. Any MCP-compatible client (Claude Code, OpenClaw, Cursor, etc.) connects and gets structured long-term memory.
+The `mcp-server/` directory contains a standalone [MCP](https://modelcontextprotocol.io) server over stdio.
+
+**What is actually verified**, as of 2026-08-24, against a live backend: protocol version
+`2025-06-18`; `initialize`, `tools/list` and `tools/call` all round-trip; `memory_recall`
+returns real results. That is a protocol-level check run directly over stdio — not a
+client-by-client compatibility matrix.
+
+Any client that speaks MCP over stdio should therefore work, but we have not sat in front
+of each one. Listed below is the config we run ourselves (Claude Code) and no others. If
+you get it working with a different client, a PR to this section is the useful kind.
 
 ### Tools
 
@@ -139,6 +148,11 @@ The `mcp-server/` directory contains a standalone [MCP](https://modelcontextprot
 
 `memory_recall` is the only one of the five that never calls a model. `memory_reflect` is an agentic loop with repeated LLM calls — if you expose this server to anything untrusted, expose `memory_recall` alone.
 
+One thing worth knowing before you budget context: the backend treats `limit` on recall as
+a retrieval hint, not a result cap — it returns everything inside its own token budget,
+around 110 facts. The MCP layer enforces your `limit` on the way out, so a `limit: 2` recall
+costs about 1 KB instead of 43 KB. Recall spends no model call; it does spend context.
+
 ### Setup
 
 ```bash
diff --git a/mcp-server/server.js b/mcp-server/server.js
index 8f4e7c6f..80abeef5 100644
--- a/mcp-server/server.js
+++ b/mcp-server/server.js
@@ -98,7 +98,10 @@ async function hindsightRequest(method, path, body = null) {
 
 const server = new McpServer({
   name: 'rcll',
-  version: '1.0.0',
+  // Keep in step with package.json and server.json — the MCP registry validates
+  // the package it resolves against this manifest, and a version that disagrees
+  // with the tarball is a review failure, not a cosmetic drift.
+  version: '0.1.0',
 });
 
 // Tool 1: memory_retain
@@ -177,7 +180,7 @@ server.tool(
 
     const result = await hindsightRequest('POST', `/${bankId}/memories/recall`, body);
 
-    const memories = (result.results || []).map(r => ({
+    const all = (result.results || []).map(r => ({
       id: r.id || r.uuid || null,
       text: r.text,
       type: r.type,
@@ -187,6 +190,13 @@ server.tool(
       hall: r.hall || null,
     }));
 
+    // The backend treats `limit` as a retrieval hint, not a result cap: it returns
+    // everything that fits its own token budget (~110 facts, ~40 KB of JSON) no
+    // matter what we ask for. Unenforced here, the tool's own description would be
+    // false and every recall would spend ~10k tokens of the caller's context.
+    const effectiveLimit = body.limit;
+    const memories = all.slice(0, effectiveLimit);
+
     return {
       content: [{
         type: 'text',
@@ -194,6 +204,9 @@ server.tool(
           success: true,
           bank_id: bankId,
           count: memories.length,
+          ...(all.length > memories.length
+            ? { returned_of_available: `${memories.length} of ${all.length}` }
+            : {}),
           memories,
         }, null, 2),
       }],