Jules Dispatch

jules-dispatch 🚀

Batch-dispatch tasks to Google Jules in parallel — and use it as an MCP tool inside Claude Code or OpenAI Codex CLI.

[npm version] [CI] [TypeScript] [MCP] [License: MIT]

🌐 Languages: English · 简体中文

[jules-dispatch banner — one orchestrator AI fanning out tasks to many parallel Jules workers, each producing a PR]

🌐 Landing page & interactive docs
https://yuuqq.github.io/jules-dispatch/ — full onboarding guide and MCP integration examples

What Is This?

jules-dispatch is a CLI and an MCP server for the Google Jules API that lets you:

• Dispatch any number of Jules tasks with up to 50 concurrent session creations, optionally paced
• Define tasks as simple YAML files — title, repo, branch, prompt
• Poll for completion and collect generated PR links
• Approve plans, send follow-up messages, cancel runaway sessions, tail live activity
• Plug into Claude Code or Codex as an MCP server — your AI assistant calls Jules as a tool

It turns Jules from a "one task at a time" tool into a massively parallel coding workforce, controlled by either humans on the CLI or other AIs over MCP.

🏗 How It Works

flowchart LR    O["🧠 Orchestrator<br/>(Claude / Codex / you)"]    O -->|writes| T["📄 tasks/*.yaml"]    T --> D["⚡ jules-dispatch batch"]    D -->|parallel| J1["🤖 Jules #1"]    D -->|parallel| J2["🤖 Jules #2"]    D -->|parallel| J3["🤖 Jules #3"]    D -->|parallel| J4["🤖 Jules #N"]    J1 --> P1["🔀 PR #1"]    J2 --> P2["🔀 PR #2"]    J3 --> P3["🔀 PR #3"]    J4 --> P4["🔀 PR #N"]
    classDef orch fill:#7c3aed,stroke:#5b21b6,color:#fff    classDef tool fill:#0891b2,stroke:#0e7490,color:#fff    classDef worker fill:#f59e0b,stroke:#b45309,color:#fff    classDef pr fill:#10b981,stroke:#047857,color:#fff    class O orch    class D tool    class J1,J2,J3,J4 worker    class P1,P2,P3,P4 pr

✨ What's New in 1.2 — Optional AI Task Planning (BYO LLM)

Entirely optional. All core commands work without any LLM key. Skip this section if you only want raw dispatch.

Stop hand-writing task YAML. Give jules-dispatch one sentence and let an LLM expand it into N parallel Jules sessions.

$ jules-dispatch auto "Migrate every Express route to Fastify and add request-validation tests"
Planning with gpt-4o-mini...
Planned 6 task(s):  1. Migrate auth routes (/api/auth/*) to Fastify  2. Migrate user routes (/api/users/*) to Fastify  3. Migrate billing routes (/api/billing/*) to Fastify  4. Replace Express middleware with Fastify hooks  5. Update server bootstrap to use Fastify instance  6. Add Vitest request-validation tests for all migrated routes
Dispatch all 6 task(s)? [y/N]

Bring your own LLM — works with any OpenAI-compatible /chat/completions endpoint:

Configure via env vars (LLM_API_KEY, LLM_BASE_URL, LLM_MODEL) or per-invocation flags (--llm-key, --llm-base-url, --llm-model). OPENAI_API_KEY and OPENROUTER_API_KEY are also recognised as fallbacks.

✨ What's New in 1.1

• 🧰 MCP server (jules-dispatch mcp) — 8 always-registered tools, plus 2 optional planning tools when an LLM key is configured, and 7 opt-in legacy aliases
• 🤖 --json mode — machine-readable output on every command for AI agents and shell pipelines
• ✅ Plan approval workflow — plan, approve commands + requirePlanApproval: true task option
• 📡 Live tailing — tail <id> streams activity events as they happen
• ❌ Cancel sessions — cancel <id> aborts runaway runs
• 🔍 Direct lookup — get <id>, status --ids no longer limited to the recent page
• 🛡️ Real failure detection — uses session.state, fails fast, distinct exit codes
• ⚡ Smart retries — exponential backoff with jitter, honours Retry-After
• 📥 Stdin input — dispatch - reads YAML/JSON from a pipe
• 🔑 --api-key flag — pass keys per-invocation, no .env required

✨ Key Features

💡 Five Common Use Cases

jules-dispatch works best when a change can be split into independent, PR-sized tasks. If tasks edit the same files or depend on earlier output, dispatch them in separate waves instead of running them concurrently.

1. Add test coverage across several modules

Suppose the auth, billing, users, and audit modules all need tests. Put one self-contained task file per module in a dedicated directory, then dispatch the directory as a batch:

jules-dispatch batch tasks/add-tests --parallel 4

Each Jules session owns one module. With AUTO_CREATE_PR enabled, the result is a set of focused PRs that can be reviewed and merged independently. A failed task can be retried without restarting the rest.

Why it helps: independent test work runs at the same time without turning into one large, hard-to-review change.

2. Break a large migration into executable tasks

For a goal such as migrating an Express API to Fastify, use the optional LLM planner to identify independent routes, middleware, startup code, and test work:

jules-dispatch auto "Migrate the Express API to Fastify and add request-validation tests" \  --max 8 --parallel 4

auto shows the proposed tasks and asks for confirmation before dispatching them. Use plan-tasks instead when you want to save and edit the generated YAML before anything is sent to Jules.

Why it helps: the planner reduces the cost of decomposing a broad goal while keeping the task boundaries visible and reviewable.

3. Roll out the same change across multiple repositories

To add a shared CI check, security baseline, or contribution policy across several Jules-connected repositories, give each task its own source:

title: "Add the security baseline to the API"prompt: "Add the agreed security checks and open a focused PR."source: "sources/github/acme/api"branch: "main"---title: "Add the security baseline to the worker"prompt: "Add the agreed security checks and open a focused PR."source: "sources/github/acme/worker"branch: "main"

Place the task file in a dedicated batch directory and dispatch it with controlled concurrency and launch pacing:

jules-dispatch batch tasks/security-baseline --parallel 6 --pace-ms 250

Why it helps: one command coordinates the rollout while preserving a separate session and PR for each repository plus an audit log for the batch.

4. Keep a human approval gate for risky changes

Authentication, authorization, and database migrations often need review before implementation begins. Require Jules to stop after planning:

title: "Refactor authorization checks"prompt: "Centralize API authorization checks without changing the public API."requirePlanApproval: true

Inspect the plan, send corrections if needed, approve it, and then continue monitoring:

jules-dispatch plan abc123jules-dispatch message abc123 "Do not change the public API"jules-dispatch plan abc123       # inspect the revised planjules-dispatch approve abc123jules-dispatch wait abc123

Why it helps: you keep control of high-impact decisions without giving up delegated execution.

5. Let Claude Code or Codex orchestrate the whole run

After configuring the MCP server, describe the outcome instead of operating each session yourself:

Analyze this repository, split its test gaps into independent tasks, and dispatch them to Jules. Ask me before approving plans or answering feedback requests. When every session finishes, summarize the outcome and PR URL for each task.

The coding assistant can call jules_dispatch, wait with jules_monitor, inspect action-required sessions with jules_interact, and return a final PR summary.

Why it helps: you manage the goal and the important decisions while the assistant handles dispatch, follow-up, and result collection.

🤖 Use Inside Claude Code or Codex (MCP)

The MCP server exposes Jules as a set of tools your coding AI can call directly.

sequenceDiagram    autonumber    actor U as 👤 You    participant CC as 💬 Claude Code / Codex    participant MCP as ⚡ jules-dispatch (MCP)    participant J as ☁️ Google Jules
    U->>CC: "Add tests to 5 modules"    CC->>MCP: jules_dispatch(tasks)    MCP->>J: POST /sessions × 5    J-->>MCP: 5 session IDs    MCP-->>CC: {dispatched: 5}
    CC->>MCP: jules_monitor(ids, wait=true)    loop until terminal or action required        MCP->>J: GET /sessions/{id}    end    MCP-->>CC: {sessions, wait: {completed, actionRequired, ...}}
    opt a session requires action        CC->>MCP: jules_interact(id)        MCP-->>CC: state + plan + activities        CC->>MCP: jules_approve_plan(id) or jules_send_message(id, text)        CC->>MCP: jules_monitor(ids, wait=true)    end
    CC->>MCP: jules_interact(id)    MCP-->>CC: terminal status + PR output    CC-->>U: "Done. PRs: #42, #43, #44, #45, #46"

Install for Claude Code

npm install -g @yuuqq/jules-dispatch

Full setup guide (including GSD integration): docs/MCP-INTEGRATION.md

Add to ~/.config/claude-code/mcp.json (or use claude mcp add):

{  "mcpServers": {    "jules-dispatch": {      "command": "jules-dispatch",      "args": ["--project", "/path/to/your/project", "mcp"],      "env": {        "JULES_API_KEY": "your-api-key-here",        "JULES_DEFAULT_SOURCE": "sources/github/owner/repo",        "JULES_DEFAULT_BRANCH": "main"      }    }  }}

Then in Claude Code: "Dispatch 5 Jules tasks to add tests to the auth, payments, users, sessions, and audit modules." Claude calls jules_dispatch, monitors them with jules_monitor, and uses jules_interact when it needs full context or PR output.

Install for OpenAI Codex CLI

Add to ~/.codex/config.toml:

[mcp_servers.jules-dispatch]command = "jules-dispatch"args = ["--project", "/path/to/your/project", "mcp"]env = { JULES_API_KEY = "your-api-key-here", JULES_DEFAULT_SOURCE = "sources/github/owner/repo" }

Install as an Agent Skill

This repository also ships a lightweight skill wrapper at skills/jules-dispatch/. The skill teaches Claude Code, Codex, or any Agent Skills-compatible host when and how to use the jules-dispatch MCP tools.

For Codex, copy or install the folder as jules-dispatch in your Codex skills directory, then restart Codex:

cp -R skills/jules-dispatch "${CODEX_HOME:-$HOME/.codex}/skills/jules-dispatch"

For Agent Skills-compatible hosts that use a shared skills directory, copy the same folder into that host's skills directory. The skill is only the instruction layer; you still need the MCP server configured with jules-dispatch mcp and a valid JULES_API_KEY.

MCP Tools Exposed

Consolidated tools (recommended)

The server registers 8 tools by default: 3 recommended consolidated tools and 5 utility tools. Two additional planning tools are registered when an LLM key is configured. Seven deprecated aliases are available behind --legacy-tools / JULES_MCP_LEGACY_TOOLS=1 but are not registered by default, so they cost no context unless you need them.

jules_dispatch — Create one or more sessions

Accepts a single task object, an array of tasks, or a YAML/JSON string.

Parameters:

{  "tasks": [    { "title": "Fix auth bug", "prompt": "Fix the null check in login()" },    { "title": "Add tests", "prompt": "Add unit tests for auth.ts" }  ],  "parallel": 5,  "paceMs": 250}

Dispatch uses a continuously replenished worker pool rather than fixed waves: whenever one task finishes creating its session, the next queued task can start. paceMs applies globally across all workers, so consecutive creation starts are separated by at least that interval while result order still matches task order.

jules_monitor — Check status or wait for the next resolution point

Parameters:

{  "sessionIds": ["abc123", "def456"],  "wait": true,  "timeoutMs": 300000}

Action-required states are AWAITING_PLAN_APPROVAL, AWAITING_USER_FEEDBACK, and PAUSED. When the result includes actionRequired, inspect those sessions with jules_interact, approve the plan or send feedback as appropriate, then call jules_monitor again for the unresolved IDs.

jules_interact — Inspect a session in full context

Returns session details, derived status, the globally latest plan, activity timeline, and PR output in one call. The server scans the complete oldest-first activity feed, then returns the newest activityCount entries in chronological order plus activityTotal for the full feed.

Parameters:

{ "sessionId": "abc123", "activityCount": 20 }

Utility tools

Optional LLM-powered tools (requires LLM key)

Response format

All tools return:

{ "success": true, "data": { ... } }

Errors return:

{  "success": false,  "error": {    "message": "Authentication failed",    "status": 401,    "name": "Error",    "recovery_hint": "Check your API key"  }}

Legacy tools (deprecated aliases, opt-in)

jules_dispatch_task, jules_dispatch_batch, jules_get_session, jules_list_activities, jules_get_plan, jules_status, and jules_wait_for_completion remain functional for compatibility but are not registered by default. Enable them with either:

jules-dispatch mcp --legacy-toolsJULES_MCP_LEGACY_TOOLS=1 jules-dispatch mcp

New integrations should use the consolidated tools.

🚀 Quick Start (Plain CLI)

Prerequisites

• Node.js 20+
• A Google Jules account and API key
• A GitHub repository connected to Jules

1. Install

npm install -g @yuuqq/jules-dispatch

2. Set up (interactive wizard)

jules-dispatch init

The wizard prompts for your API key, default source, and branch. It writes a .env file.

For CI/scripts (non-interactive):

jules-dispatch init --api-key sk-xxx --source sources/github/owner/repo

3. Validate your setup

jules-dispatch doctor

4. Write a task

# tasks/add-dark-mode.yamltitle: "Add Dark Mode Support"prompt: |  Add a dark mode toggle to the React app:  1. Add a ThemeContext with light/dark state  2. Wrap App with ThemeProvider  3. Add a toggle button in the Header  4. Persist preference in localStorage  5. Open a PR

5. Dispatch it

jules-dispatch dispatch tasks/add-dark-mode.yaml# ✓ Add Dark Mode Support#   Session: https://jules.google.com/session/abc123#   ID:      abc123

6. Batch-dispatch a whole directory

jules-dispatch batch tasks/ --parallel 10

That's it — 6 steps from install to your first PR.

📖 CLI Reference

Global flags

Commands

Exit codes (for shell scripts and agents)

Monitoring error behavior

status reports session or activity lookup failures explicitly as status: "error" and exits nonzero instead of fabricating a Jules failure or trusting a potentially stale state. Polling commands retry only network, rate-limit, and server errors; invalid requests, authentication failures, and missing sessions fail immediately with the affected session ID in the error context.

dispatch examples

# Override repo/branchjules-dispatch dispatch tasks/my-task.yaml \  --source sources/github/org/other-repo --branch develop
# Read from stdinecho 'title: Quick fix\nprompt: Fix typo in README' | jules-dispatch dispatch -
# JSON output (great for piping)jules-dispatch dispatch tasks/my-task.yaml --json | jq -r '.sessionId'

batch examples

jules-dispatch batch tasks/                       # default tasks/ dirjules-dispatch batch tasks/ --parallel 20         # 20 concurrentjules-dispatch batch tasks/ --parallel 10 --pace-ms 250  # globally space starts by 250 msjules-dispatch batch tasks/ --no-log              # don't write dispatch logjules-dispatch batch tasks/ --json                # one JSON summary at the end

batch and auto both use the same continuously replenished worker pool. --parallel caps in-flight session creation and --pace-ms sets the global minimum spacing between creation starts; it does not add a delay separately inside each worker.

wait example

# Chain dispatch → wait via JSON output:ID=$(jules-dispatch dispatch tasks/x.yaml --json | jq -r '.sessionId')jules-dispatch wait "$ID" --interval 10000 --timeout 1800000

tail example

jules-dispatch tail abc123                        # human-readable streamjules-dispatch tail abc123 --json                 # NDJSON event stream

🔄 Session Lifecycle

jules-dispatch tracks every Jules session through its full lifecycle and surfaces each state through the CLI / MCP:

stateDiagram-v2    [*] --> QUEUED    QUEUED --> PLANNING    PLANNING --> IN_PROGRESS    PLANNING --> AWAITING_PLAN_APPROVAL: approval required    AWAITING_PLAN_APPROVAL --> IN_PROGRESS: approve plan    IN_PROGRESS --> AWAITING_USER_FEEDBACK: input required    AWAITING_USER_FEEDBACK --> IN_PROGRESS: send feedback    IN_PROGRESS --> PAUSED: execution paused    PAUSED --> IN_PROGRESS: execution resumes    IN_PROGRESS --> COMPLETED: success    IN_PROGRESS --> FAILED: error    COMPLETED --> [*]    FAILED --> [*]

For compatibility, jules-dispatch also normalizes legacy API states: PENDING, RUNNING, AWAITING_USER_INPUT, CANCELLED, and CANCELED. Cancellation is sent with the Jules DELETE /sessions/{id} operation.

📄 Task File Format

Field Reference

YAML example

title: "Add unit tests for auth module"prompt: |  Add comprehensive unit tests for src/auth.ts:  1. Test login with valid credentials  2. Test login with invalid credentials  3. Test token refresh flow  4. Test session expiry handling  5. Open a PR with the test filesource: "sources/github/myorg/myrepo"branch: "develop"autoMode: "AUTO_CREATE_PR"requirePlanApproval: false

Multiple tasks in one file (YAML --- separators)

title: "Fix lint errors in src/auth"prompt: "Fix all ESLint errors in src/auth.ts"---title: "Fix lint errors in src/api"prompt: "Fix all ESLint errors in src/api.ts"---title: "Fix lint errors in src/utils"prompt: "Fix all ESLint errors in src/utils.ts"

JSON format

{  "title": "Fix the thing",  "prompt": "Find the bug in src/auth.ts and fix it.",  "source": "sources/github/owner/repo",  "branch": "main"}

JSON array (for batch dispatch via MCP):

[  { "title": "Task 1", "prompt": "Do thing A" },  { "title": "Task 2", "prompt": "Do thing B" }]

🤖 AI-Orchestrated Parallel Development

The killer use case: combine jules-dispatch with Claude Code or Codex.

"I have a Node.js backend that needs to be migrated from Express to Fastify. Analyse the codebase, split the work into independent migration units, and dispatch them all to Jules in parallel using the jules-dispatch MCP tools. Then poll for completion and report back the PR URLs."

With the MCP server installed, your assistant will:

1. Analyse your codebase
2. Commit and push the target branch, because Jules works from the remote source branch rather than unpushed local changes
3. Call jules_dispatch with N task definitions
4. Call jules_monitor with wait: true; if action is required, inspect with jules_interact, approve or send feedback, and monitor again
5. Use jules_interact to collect terminal context and PR URLs

You get N parallel coding agents orchestrated by one strategic agent, hands-free.

📁 Project Structure

jules-dispatch/├── src/│   ├── cli.ts          CLI entry point (Commander)│   ├── client.ts       Jules REST client (retries, pagination)│   ├── config.ts       .env + task file loading│   ├── dispatcher.ts   Task dispatch logic│   ├── collector.ts    Status polling & wait│   ├── errors.ts       Structured error translation (Problem/Cause/Fix)│   ├── output.ts       Text vs JSON output mode, color detection│   ├── init.ts         Interactive init wizard│   ├── doctor.ts       Environment validation (doctor command)│   ├── mcp.ts          MCP server (8 tools + 2 optional planner tools + 7 opt-in aliases)│   ├── mcp-helpers.ts  MCP response helpers (ok/fail/recovery hints)│   ├── polling.ts      Shared poll-with-callback engine│   ├── tail.ts         Bounded, cursor-aware activity tailing│   ├── planner.ts      Optional LLM task planner│   ├── log.ts          Verbose logging│   └── types.ts        TypeScript types├── tasks/              Your task YAMLs live here├── .env                Generated by jules-dispatch init└── .dispatch-logs/     JSON audit trail

🛠 Development

npm installnpm run build     # compile TypeScript → dist/npm run dev       # run CLI directly with tsxnpm run lintnpm run test

📜 License

MIT — see LICENSE

Built to make Google Jules actually scale.