Home/ Knowledge Base/ Integrations & AI/ Dispatch — the built-in assistant
Integrations & AI

Dispatch — the built-in assistant

Ask questions in plain English and get answers built on your own lanes, recipes and servers — plus release notes, version-bump suggestions, failure diagnosis and recipe drafting. Powered by your own Claude Code, with your secrets kept out of it.

Last updated 2026-07-21

Dispatch is AppRelay's built-in assistant. Ask it anything about shipping your project — in plain English — and it answers using your actual setup: your lanes, your recipes, your servers, your toolchains. Ask it to set something up and it doesn't just describe the steps, it proposes the config as a card you review and accept.

Powered by your own Claude Code. AppRelay shells out to the claude command-line tool you already have installed and signed in. There's no AppRelay AI service and no API key to paste into AppRelay — it uses your existing Claude account, and your Anthropic credentials never pass through AppRelay.

What it's like to use

Open Dispatch in the sidebar and you get a prompt box. Ask something like:

“How do I set up an FTP lane that uploads only staged or newly edited files to my server?”

You get back the click-path for your app (Settings → Servers → Add server, then Project Settings → Lanes → Distributions → Add), the right built-in recipe for the job, and — because you asked it to set something up — a proposal card holding the actual server entry and recipe steps. Press Add and it lands in your workspace. Press nothing and nothing happens.

Dispatch reads, it never writes. It runs with read-only tools, so it can open your project to answer accurately but cannot edit a file or run a command. The only way it changes anything is a proposal you explicitly accept.

What it knows

Your live configuration
Every question carries the selected project's lanes, toolchains, version sources, distributions, test suites, and your saved servers and custom recipes — so answers name your real lanes instead of inventing examples.
Your repository
It can read, search and list files in the project folder to check a detail before answering.
Your CLAUDE.md
Because it runs inside the repo, Claude Code loads the project's CLAUDE.md automatically. The composer shows a chip when it's in play.
How AppRelay itself works
A version-matched guide to the app you're running — navigation, concepts, task walkthroughs — so it gives real click-paths rather than guesses. It updates independently of app releases.
Saved plans (opt-in)
Plans under ~/.claude/plans are never sent automatically. Attach one with the paperclip to include it as background for your next question.

The tabs

Chat
The conversation. Answers stream in, and history is kept per project — switching projects swaps to that project's threads.
Context
Browse the Claude Code files this repo ships: CLAUDE.md, TASKS.md, slash-commands and subagents.
Plans
Your saved plans, with pin-to-project and attach-to-chat.
Usage
A token ledger: input, output and cache tokens per question, totals, and estimated cost. On a Claude subscription the questions are covered by your plan, so the cost column is the equivalent API price shown for scale.

The header shows which Claude account is signed in and on what plan, so it's obvious when a sign-in has expired — and a project picker to switch which project you're asking about.

The rest of the AI actions

Beyond Dispatch, the same engine powers focused actions where the work happens:

Generate release notes
Reads your commits since the last tag and drafts clean, user-facing notes grouped into Features / Fixes / Other.
Suggest a version bump
Looks at the same commits and recommends major, minor, patch, or none — advisory context next to the manual bump buttons.
Explain & fix a failure
When a build, test, or distribution step fails, it reads the error and suggests a likely cause, a concrete fix, and how to verify it.
Draft recipe steps
Describe what you want in plain English and get real shell steps to review before they're added.
Review a recipe
Lints a recipe for wrong flags, missing signing, hardcoded values that should be {placeholders}, and things that would fail on a clean machine.

Before you start

  1. Install Claude Code. AppRelay looks for the claude binary on your system. If it isn't found, Dispatch shows an Install Claude Code link instead.
  2. Sign in. If you're not signed in, the account chip in the Dispatch header says so and offers Sign in, which opens a terminal running claude auth login. Come back and press Recheck.

If a sign-in expires while you're working, the chip turns red and says Session expired — signing in again clears it.

What needs Studio

Two actions work on the free tier as long as Claude Code is installed; the rest are part of AppRelay Studio:

FeatureAvailable on
Generate release notesFree + Studio
Suggest a version bumpFree + Studio
Dispatch (chat, proposals, usage)Studio
Explain & fix a failureStudio
Draft recipe stepsStudio
Review a recipeStudio

All of them require Claude Code to be installed and signed in — that's separate from your AppRelay license.

What leaves your machine

Because these features run through your claude tool, any context goes to Anthropic under your own account — the same as if you'd run claude in that repo yourself. AppRelay keeps what it sends deliberately narrow:

  • Dispatch: your question, a description of AppRelay's model and navigation, and a summary of the selected project's configuration — lane labels, toolchain ids, recipe references, server host/port/username. It can also read repo files to answer, exactly as the CLI would.
  • Release notes / version bump: the list of commit subjects since the last tag.
  • Explain & fix: the failed command, a capped tail of its error output, the working directory, and the toolchain name.
  • Recipe drafting / review: the recipe's kind, target toolchain, your description, and the available placeholder names.

Your secrets and signing credentials are never included — server passwords stay in your Keychain (macOS) or encrypted vault (Windows) and are referenced, never sent. Conversations and their token counts are stored locally on your machine only. See Security & privacy for how this fits the local-first model.

Beyond the window

The same engine is exposed two other ways for automation: an MCP server, so an agent can list your projects, read versions and git status, and run builds, tests, and releases; and a headless apprelay CLI (build, run, test, release) that reads the same workspace and prints JSON for your own scripts and CI. The irreversible release action is always gated behind an explicit confirmation.

Next steps