Re-add MCP bridge so Hermes can delegate to the real Claude Code CLI

Recreates the PR #17 change (closed without merging, then asked for back
after confirming a real ANTHROPIC_API_KEY would mean separate/duplicate
billing rather than actually using the Pro/Max subscription).

New POST /mcp endpoint (Streamable HTTP transport, stateless — fresh
McpServer+transport per request) exposing one tool, ask_claude_code: runs
the real `claude` binary against a prompt, billed against the subscription
rather than API credits. Works specifically because it's the actual CLI
making the request server-side. Read-only — no Edit/Write/git-push/
git-commit tools.

Also removes litellm-config.yaml's "anthropic-claude" model entry: it only
ever worked when the real claude CLI itself was the caller (proven earlier),
so having it listed as a selectable model was actively misleading — Hermes
picking it directly is exactly what produced the '401: Missing Anthropic
API Key' confusion that led back to this bridge. The MCP tool is the actual
working path now; general_settings.forward_client_headers_to_llm_api is
also removed since nothing uses it anymore.

Tested end-to-end locally again before pushing (built the image, ran it,
full MCP handshake via curl) — real 'pong' from the real claude CLI.

.env.example

@@ -21,6 +21,12 @@ GITEA_REGISTRY_IMAGE=gitea.apps.williamturner.eu/<your-gitea-username>/<repo-nam
 # Run `claude setup-token` interactively (needs a browser + Claude Pro/Max subscription)
 # to generate this — it's a long-lived OAuth token, not an API key.
 CLAUDE_CODE_OAUTH_TOKEN=
+# Any random string — shared secret for claude-agent's /mcp bridge endpoint (see
+# agent/src/mcpBridge.js), which lets Hermes delegate a question to the real `claude`
+# CLI (billed against the subscription above) via MCP. Register it in Hermes with:
+#   docker exec hermes hermes config set mcp_servers.claude-code.url http://claude-agent:3001/mcp
+#   docker exec hermes hermes config set 'mcp_servers.claude-code.headers.Authorization' 'Bearer <this value>'
+MCP_BRIDGE_KEY=
 
 # --- litellm (local LLM gateway — used by Hermes, see litellm-config.yaml) ---
 OPENROUTER_API_KEY=

agent/package.json

@@ -8,6 +8,8 @@
     "start": "node src/server.js"
   },
   "dependencies": {
-    "express": "^4.19.2"
+    "express": "^4.19.2",
+    "@modelcontextprotocol/sdk": "^1.30.0",
+    "zod": "^3.23.8"
   }
 }

agent/src/mcpBridge.js

@@ -0,0 +1,71 @@
+import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
+import { z } from "zod";
+import { execFile } from "node:child_process";
+import { promisify } from "node:util";
+import { mkdtemp, rm, mkdir } from "node:fs/promises";
+import path from "node:path";
+
+const execFileAsync = promisify(execFile);
+const WORKSPACE_ROOT = "/workspace";
+
+// Runs the real Claude Code CLI — billed against the Claude Pro/Max subscription
+// (CLAUDE_CODE_OAUTH_TOKEN), not per-token API billing. This only works because it's
+// the actual `claude` binary making the request: Anthropic rejects the same OAuth token
+// used by any other HTTP client (proven earlier — direct curl replicating the same
+// request shape gets rejected). Read-only: no git/file-write tools, since this is a
+// quick-answer bridge, not a repo-editing agent (claude-agent's own webhook flow already
+// owns that for PRs).
+async function askClaudeSubscription(prompt) {
+  await mkdir(WORKSPACE_ROOT, { recursive: true });
+  const dir = await mkdtemp(path.join(WORKSPACE_ROOT, "mcp-"));
+  try {
+    const { stdout } = await execFileAsync(
+      "claude",
+      [
+        "-p", prompt,
+        "--output-format", "text",
+        "--permission-mode", "bypassPermissions",
+        "--disallowedTools", "Bash(git push:*),Bash(git commit:*),Edit,Write,NotebookEdit",
+      ],
+      { cwd: dir, maxBuffer: 1024 * 1024 * 32 }
+    );
+    return stdout;
+  } finally {
+    await rm(dir, { recursive: true, force: true });
+  }
+}
+
+// A fresh McpServer per request (stateless transport) — cheap, and avoids any
+// cross-request state for what's a single-tool, single-shot bridge.
+export function createMcpServer() {
+  const server = new McpServer({ name: "claude-code-bridge", version: "1.0.0" });
+
+  server.registerTool(
+    "ask_claude_code",
+    {
+      description:
+        "Ask the real Claude Code CLI a question or reasoning task, billed against the " +
+        "Claude Pro/Max subscription rather than per-token API credits. Use this when " +
+        "you specifically want Claude's own model rather than whatever the default " +
+        "routed model provides. Read-only — cannot edit files, push, or commit.",
+      inputSchema: { prompt: z.string().describe("The question or task to ask Claude") },
+    },
+    async ({ prompt }) => {
+      try {
+        const text = await askClaudeSubscription(prompt);
+        return { content: [{ type: "text", text }] };
+      } catch (err) {
+        return { content: [{ type: "text", text: `Error: ${err.message}` }], isError: true };
+      }
+    }
+  );
+
+  return server;
+}
+
+export function mcpAuthMiddleware(req, res, next) {
+  const key = process.env.MCP_BRIDGE_KEY;
+  if (!key) return res.status(500).send("MCP_BRIDGE_KEY not configured");
+  if (req.get("Authorization") !== `Bearer ${key}`) return res.status(401).send("unauthorized");
+  next();
+}

agent/src/server.js

@@ -1,7 +1,9 @@
 import express from "express";
 import crypto from "node:crypto";
+import { StreamableHTTPServerTransport } from "@modelcontextprotocol/sdk/server/streamableHttp.js";
 import { postPRComment } from "./gitea.js";
 import { reviewPullRequest } from "./runner.js";
+import { createMcpServer, mcpAuthMiddleware } from "./mcpBridge.js";
 
 const app = express();
 app.use(
@@ -56,6 +58,33 @@ app.post("/webhooks/gitea", async (req, res) => {
   }
 });
 
+// MCP bridge — lets Hermes (or anything else speaking MCP) delegate a question to the
+// real Claude Code CLI, billed against the subscription. Stateless: a fresh server +
+// transport per request, no session tracking needed for a single-tool bridge like this.
+app.post("/mcp", mcpAuthMiddleware, async (req, res) => {
+  const mcpServer = createMcpServer();
+  const transport = new StreamableHTTPServerTransport({ sessionIdGenerator: undefined });
+  res.on("close", () => {
+    transport.close();
+    mcpServer.close();
+  });
+  try {
+    await mcpServer.connect(transport);
+    await transport.handleRequest(req, res, req.body);
+  } catch (err) {
+    console.error("MCP request handling failed:", err);
+    if (!res.headersSent) res.status(500).send("internal error");
+  }
+});
+
+app.get("/mcp", mcpAuthMiddleware, (_req, res) => {
+  res.status(405).set("Allow", "POST").send("Method Not Allowed");
+});
+
+app.delete("/mcp", mcpAuthMiddleware, (_req, res) => {
+  res.status(405).set("Allow", "POST").send("Method Not Allowed");
+});
+
 app.listen(PORT, () => {
   console.log(`claude-agent listening on :${PORT}`);
 });

docker-compose.yml

@@ -120,9 +120,10 @@ services:
       - "traefik.http.services.hermes-dashboard.loadbalancer.server.port=9119"
 
   claude-agent:
-    # Gitea PR-review only now — no Matrix presence (see hermes above; only one agent
-    # is meant to be in Matrix). Still triggered by Gitea's pull_request webhook and
-    # posts review comments there, entirely independent of Matrix/LiteLLM.
+    # No Matrix presence (see hermes above; only one agent is meant to be in Matrix).
+    # Two things call this now: Gitea's pull_request webhook (PR review), and Hermes,
+    # over MCP (POST /mcp), to delegate a question to the real `claude` CLI when it
+    # specifically wants the Claude subscription instead of whatever LiteLLM routed it to.
     image: ${GITEA_REGISTRY_IMAGE}
     container_name: claude-agent
     restart: unless-stopped
@@ -136,6 +137,9 @@ services:
       # Claude subscription (Pro/Max) auth via `claude setup-token`, not API billing —
       # Claude Code reads this in preference to ANTHROPIC_API_KEY when both could apply.
       CLAUDE_CODE_OAUTH_TOKEN: ${CLAUDE_CODE_OAUTH_TOKEN}
+      # Shared secret for the /mcp bridge endpoint (internal network only either way, but
+      # this keeps it from being callable by anything that merely reaches the container).
+      MCP_BRIDGE_KEY: ${MCP_BRIDGE_KEY}
     volumes:
       - agent_workspace:/workspace
     networks:

litellm-config.yaml

@@ -11,18 +11,15 @@ model_list:
       model: openrouter/openai/gpt-4o-mini
       api_key: os.environ/OPENROUTER_API_KEY
 
-  # Routes to Anthropic using the CALLER's forwarded Authorization header (the Claude
-  # Pro/Max subscription OAuth token) instead of a LiteLLM-held API key — billed against
-  # the subscription, not per-token. CONFIRMED WORKING, but only for the real `claude`
-  # CLI binary as caller (tested: `claude -p` with ANTHROPIC_BASE_URL pointed here
-  # returned a real completion). An earlier test with plain curl replicating the same
-  # request shape failed — Anthropic apparently requires header/fingerprint details only
-  # the real CLI sends, which LiteLLM faithfully relays but a hand-built request won't
-  # have. Do NOT expect this to work for other callers (Hermes, generic HTTP clients) —
-  # they aren't the real CLI and can't reproduce that fingerprint.
-  - model_name: anthropic-claude
-    litellm_params:
-      model: anthropic/claude-sonnet-5
+  # NOT included: an "anthropic-claude" model routing to Anthropic via the caller's
+  # forwarded OAuth header (general_settings.forward_client_headers_to_llm_api). It
+  # genuinely works — but ONLY when the real `claude` CLI binary is the caller (its
+  # request carries a header/fingerprint only that binary sends; a hand-built request,
+  # including Hermes selecting this model directly, gets a hard auth error from
+  # Anthropic). Having it selectable here caused exactly that confusion once already.
+  # The actual working path for "Hermes uses the Claude subscription" is the MCP bridge
+  # at claude-agent's /mcp (agent/src/mcpBridge.js) — it shells out to the real `claude`
+  # binary server-side instead of trying to make an arbitrary caller impersonate it.
 
 litellm_settings:
   # Callers (Hermes included) send provider-specific params like reasoning_effort that
@@ -30,5 +27,4 @@ litellm_settings:
   drop_params: true
 
 general_settings:
-  forward_client_headers_to_llm_api: true
   master_key: os.environ/LITELLM_MASTER_KEY