Code Mode

Write and execute Bun scripts that call MCP tools programmatically — multiple calls on a single persistent connection, without returning to the LLM between each call.

When to use

• Multi-tool orchestration: 3+ MCP tool calls where intermediate results feed into subsequent calls
• Batch operations: Repeated calls with different parameters (e.g., process a list of items)
• Cross-server workflows: Combining results from different MCP servers
• Token-heavy tool sets: Servers with many tools (like figma-console's 63+) where loading all schemas wastes context

When NOT to use

• Simple single-tool calls → use the normal tool-call loop
• Tools that require interactive OAuth → authenticate via claude mcp first, then use code-mode
• Operations that need human approval between steps → use normal tool-call loop

---

The Pattern

Scripts are written to /tmp/claude-code-mode/ (auto-cleaned on session end via hook). The SDK resolves via NODE_PATH pointing to the skill's node_modules/.

Step 1: Write the script

mkdir -p /tmp/claude-code-mode

Write the script to /tmp/claude-code-mode/_run.ts:

#!/usr/bin/env bun
import { getServerConfig, createTransport } from "<path-to-skill>/scripts/mcp-client.ts";
import { Client } from "@modelcontextprotocol/sdk/client/index.js";

const config = getServerConfig("SERVER_NAME");
if (!config) { console.error("Server not found"); process.exit(1); }

const client = new Client({ name: "cc-code-mode", version: "1.0.0" });
const transport = await createTransport(config);

try {
  await client.connect(transport);

  // Call tools — connection stays open between calls
  const result1 = await client.callTool({
    name: "tool_name",
    arguments: { key: "value" },
  });
  if (result1.isError) throw new Error(JSON.stringify(result1.content));

  const result2 = await client.callTool({
    name: "another_tool",
    arguments: { data: result1.content },
  });

  // Output structured JSON for Claude Code to parse
  console.log(JSON.stringify({ result1: result1.content, result2: result2.content }));
} catch (err) {
  console.error(JSON.stringify({ error: String(err) }));
  process.exit(1);
} finally {
  await client.close();
}

Step 2: Execute the script

NODE_PATH=<path-to-skill>/scripts/node_modules bun /tmp/claude-code-mode/_run.ts

Step 3: Parse and present results

Read stdout (JSON results) and stderr (server logs/errors). Present findings to the user.

---

Helper API

The helper at ./mcp-client.ts reads Claude Code's own config files — no duplicate configuration needed.

getServerConfig(name: string): ServerConfig | null

Returns the resolved config for a named MCP server. Reads from all Claude Code config sources (plugins, user, project) and merges by precedence. Env vars are resolved (${VAR}, ${VAR:-default}).

Returns null if the server is not configured.

listServers(): { name, type, source }[]

Returns all available MCP servers with their transport type and config source.

createTransport(config: ServerConfig): Transport

Creates the appropriate MCP transport (StdioClientTransport, StreamableHTTPClientTransport, or SSEClientTransport) from a resolved config. For stdio servers, env vars from the config are explicitly passed to the child process (the SDK filters env vars by default — this is handled automatically).

---

Available servers

Before writing a script, check what's available:

bun <path-to-skill>/scripts/mcp-client.ts list

To inspect a specific server's resolved config:

bun <path-to-skill>/scripts/mcp-client.ts get figma-console

---

Discovering tools

If you don't know a server's tools, use listTools():

await client.connect(transport);
const { tools } = await client.listTools();
console.log(JSON.stringify(tools.map(t => ({
  name: t.name,
  description: t.description,
  inputSchema: t.inputSchema,
}))));

---

Critical rules

1. Always use try/finally with client.close(). MCP servers (especially stdio) are child processes. Without close(), they become orphans.

2. Always check isError. Tool errors are returned as { isError: true, content: [...] }, NOT thrown. An unchecked error silently produces wrong results.

3. Output JSON to stdout. Claude Code reads the script's stdout. Use console.log(JSON.stringify(...)) for results. Use console.error() for debugging — it goes to stderr.

4. The connection is persistent. After client.connect(), you can call callTool() as many times as needed. No reconnection between calls.

5. Default timeout is 60s. For slow tools, pass options: client.callTool({ name, arguments }, { timeout: 120_000 }).

---

Error handling

Load: references/error-handling.md for detailed patterns.

Quick reference:

// Tool-level error (returned, not thrown)
const result = await client.callTool({ name: "...", arguments: {} });
if (result.isError) {
  console.error("Tool error:", JSON.stringify(result.content));
  // Decide: skip, retry, or abort
}

// Connection/protocol error (thrown)
try {
  await client.connect(transport);
} catch (err) {
  console.error("Connection failed:", err);
  process.exit(1);
}

// Timeout
try {
  const result = await client.callTool(
    { name: "slow_tool", arguments: {} },
    { timeout: 120_000 },
  );
} catch (err) {
  if (err.message?.includes("timed out")) {
    console.error("Tool timed out");
  }
}

---

Transport types

All three MCP transport types are supported. createTransport() auto-selects based on the config's type field.

Load: references/transport-patterns.md for details on each transport.

Config type	Transport	Notes
stdio	StdioClientTransport	Local process. Env vars passed explicitly.
http	StreamableHTTPClientTransport	Remote server. Headers supported.
sse	SSEClientTransport	Legacy. Same API as http.

---

Examples

Load: references/examples.md for full worked examples.

Multi-tool Figma workflow

import { getServerConfig, createTransport } from "<path-to-skill>/scripts/mcp-client.ts";
import { Client } from "@modelcontextprotocol/sdk/client/index.js";

const config = getServerConfig("figma-console");
if (!config) { console.error("figma-console not configured"); process.exit(1); }

const client = new Client({ name: "cc-code-mode", version: "1.0.0" });
const transport = await createTransport(config);

try {
  await client.connect(transport);

  const tokens = await client.callTool({
    name: "figma_get_variables",
    arguments: { fileUrl: "https://figma.com/design/FILE/...", format: "css" },
  });

  const fileData = await client.callTool({
    name: "figma_get_file_data",
    arguments: { fileUrl: "https://figma.com/design/FILE/...", depth: 1 },
  });

  console.log(JSON.stringify({ tokens: tokens.content, fileData: fileData.content }));
} finally {
  await client.close();
}

Cross-server workflow

import { getServerConfig, createTransport } from "<path-to-skill>/scripts/mcp-client.ts";
import { Client } from "@modelcontextprotocol/sdk/client/index.js";

// Connect to server A
const configA = getServerConfig("server-a");
const clientA = new Client({ name: "cc-code-mode", version: "1.0.0" });
const transportA = await createTransport(configA!);

// Connect to server B
const configB = getServerConfig("server-b");
const clientB = new Client({ name: "cc-code-mode", version: "1.0.0" });
const transportB = await createTransport(configB!);

try {
  await clientA.connect(transportA);
  await clientB.connect(transportB);

  const dataA = await clientA.callTool({ name: "get_data", arguments: {} });
  const dataB = await clientB.callTool({
    name: "process_data",
    arguments: { input: dataA.content },
  });

  console.log(JSON.stringify({ dataA: dataA.content, dataB: dataB.content }));
} finally {
  await clientA.close();
  await clientB.close();
}