diff --git a/package.json b/package.json
index 80b1b99..717d910 100644
--- a/package.json
+++ b/package.json
@@ -1,6 +1,6 @@
 {
   "name": "agenthub",
-  "version": "0.8.1",
+  "version": "0.9.0",
   "description": "Local coordination layer for AI coding agents",
   "type": "module",
   "main": "./dist/index.js",
diff --git a/src/cli/commands/agentSetup.ts b/src/cli/commands/agentSetup.ts
new file mode 100644
index 0000000..f407f6f
--- /dev/null
+++ b/src/cli/commands/agentSetup.ts
@@ -0,0 +1,65 @@
+import { readFileSync, writeFileSync, existsSync, mkdirSync } from 'node:fs';
+import { join } from 'node:path';
+import { homedir } from 'node:os';
+import { findProjectRoot } from '../../core/paths.js';
+
+/**
+ * Auto-start: make a Claude Code agent enter the agenthub work-loop the moment a
+ * session starts — no manual "call agenthub_work" needed.
+ *
+ * Mechanism: a verified Claude Code `SessionStart` hook (.claude/settings.json)
+ * whose command runs `agenthub hook-context`, which prints the work-loop
+ * instruction. That text is injected into the new session as context.
+ *
+ * Honest limit (per the Claude Code docs): SessionStart context is a strong
+ * NUDGE the model reads, not a hard directive it must obey — so the guide
+ * reinforces it. It also can't block startup. Codex/Kimi aren't Claude Code;
+ * their guide (CODEX.md / KIMI.md) carries the same "call agenthub_work first"
+ * nudge.
+ */
+
+/** Printed by the SessionStart hook → injected as session context. */
+export function hookContext(opts: { agent: string; role: string }): void {
+  const { agent, role } = opts;
+  process.stdout.write(
+    `AgentHub auto-start — you are "${agent}" (${role}). Before anything else, run:\n` +
+      `  agenthub work --agent ${agent} --role ${role}\n` +
+      `(or call the agenthub_work tool with { agent: "${agent}", role: "${role}" }), then LOOP: ` +
+      `work -> implement -> agenthub_task_review -> work. agenthub_work claims your tasks AND ` +
+      `surfaces your messages. Never run agenthub_task_done. Keep looping so you stay reachable.\n`,
+  );
+}
+
+export function agentSetup(cwd: string, opts: { agent: string; role: string; user?: boolean }): void {
+  const base = opts.user ? homedir() : (findProjectRoot(cwd) ?? cwd);
+  const dir = join(base, '.claude');
+  const settingsPath = join(dir, 'settings.json');
+  const command = `agenthub hook-context --agent ${opts.agent} --role ${opts.role}`;
+  const hookGroup = { matcher: 'startup|resume', hooks: [{ type: 'command', command }] };
+
+  let settings: { hooks?: { SessionStart?: unknown[] } } = {};
+  if (existsSync(settingsPath)) {
+    try {
+      settings = JSON.parse(readFileSync(settingsPath, 'utf-8')) as typeof settings;
+    } catch {
+      settings = {};
+    }
+  }
+  if (!settings.hooks) settings.hooks = {};
+  if (!Array.isArray(settings.hooks.SessionStart)) settings.hooks.SessionStart = [];
+
+  // Idempotent: drop any prior agenthub hook (identified by the command), re-add.
+  settings.hooks.SessionStart = (settings.hooks.SessionStart as Array<{ hooks?: Array<{ command?: string }> }>).filter(
+    (g) => !(g?.hooks ?? []).some((h) => typeof h?.command === 'string' && h.command.includes('agenthub hook-context')),
+  );
+  (settings.hooks.SessionStart as unknown[]).push(hookGroup);
+
+  mkdirSync(dir, { recursive: true });
+  writeFileSync(settingsPath, JSON.stringify(settings, null, 2) + '\n', 'utf-8');
+
+  const out = (s: string) => process.stdout.write(s + '\n');
+  out(`AgentHub: auto-start hook written -> ${settingsPath}`);
+  out(`  New Claude Code sessions ${opts.user ? '(all projects)' : 'in this project'} start as "${opts.agent}" (${opts.role}) and enter the work loop.`);
+  out('  Note: SessionStart context is a strong nudge, not a hard guarantee — the guide reinforces it.');
+  out('  Codex/Kimi are not Claude Code: their guide already nudges agenthub_work on first turn.');
+}
diff --git a/src/cli/index.ts b/src/cli/index.ts
index 28854d1..3482e20 100644
--- a/src/cli/index.ts
+++ b/src/cli/index.ts
@@ -6,6 +6,7 @@ import { taskCreate, taskList, taskShow, taskClaim, taskDone, taskReview, taskRe
 import { handoffCreate, handoffRead, handoffList } from './commands/handoff.js';
 import { decisionCreate, decisionList } from './commands/decision.js';
 import { messageSend, inboxList } from './commands/message.js';
+import { agentSetup, hookContext } from './commands/agentSetup.js';
 import { delegate } from './commands/delegate.js';
 import { serverStart } from './commands/server.js';
 import { update } from './commands/update.js';
@@ -117,7 +118,7 @@ async function runRemote(serverUrl: string, fn: () => Promise<void>): Promise<vo
 export function createProgram(cwd: string): Command {
   const program = new Command('agenthub')
     .description('Local coordination layer for AI coding agents')
-    .version('0.8.1')
+    .version('0.9.0')
     .option('--server <url>', 'AgentHub server URL (env: AGENTHUB_SERVER)');
 
   program
@@ -493,6 +494,29 @@ export function createProgram(cwd: string): Command {
       }
     });
 
+  // ─── auto-start ──────────────────────────────────────────────────────────
+  const agentCmd = new Command('agent').description('Per-agent machine setup');
+  agentCmd
+    .command('setup')
+    .description('Auto-start: write a Claude Code SessionStart hook so sessions enter the work loop')
+    .requiredOption('--agent <name>', 'This machine\'s agent name')
+    .requiredOption('--role <role>', 'This agent\'s role (implementer / architect / tester)')
+    .option('--user', 'Install for all projects (~/.claude) instead of just this one')
+    .action((options: { agent: string; role: string; user?: boolean }) => {
+      agentSetup(cwd, { agent: options.agent, role: options.role, user: options.user });
+    });
+  program.addCommand(agentCmd);
+
+  // Called by the SessionStart hook; prints the work-loop instruction as context.
+  program
+    .command('hook-context')
+    .description('Print the auto-start work-loop instruction (used by the SessionStart hook)')
+    .requiredOption('--agent <name>', 'Agent name')
+    .requiredOption('--role <role>', 'Agent role')
+    .action((options: { agent: string; role: string }) => {
+      hookContext({ agent: options.agent, role: options.role });
+    });
+
   program
     .command('delegate')
     .description('Suggest or auto-delegate open tasks')