loadout · in development
Every chat starts with onlywhat this project uses.
One public bug report shows Claude Code at about 56,800 tokens after someone typed “hi”. loadout learns what each project uses from the chat history on your computer and starts every new chat there with only that. Anything it cut is one command away.
- Works with
- Claude Code, Codex
- Runs on
- Mac and Linux; Windows at v1.0
- Status
- Not built yet. The audit comes first.
What loads before you type
Every new chat starts with this pile already in it. Most of it is stuff you added: skills, agents, connectors, plugins. Not every project needs all of it. Pick a plan or flip the switches and watch the total.
Example numbers to show how it works, not a measurement or a promise. Your real numbers come from your own chat history, in the first release. A plan applies from the next chat you open.
Every chat starts with all of it.
Under 8 chats of history: Claude skills this project hasn't used are shortened to their names. Nothing is switched off.
What this project used lately loads in full. The rest is shortened to its name or switched off for this chat. Safety plugins and instruction files always stay.
Your own mix.
Before you type: 52,400 tokens.
- Claude's built-in instructions and tools13,400 tokens · Can't be switched off
- Instruction files (CLAUDE.md, memory)2,600 tokens · Always sent. loadout reports on these, never cuts them
Measured, not guessed
Claude Code and Codex already keep a history of every chat on your computer, including how many tokens each request used. loadout reads it. So the audit works on day one, before you set anything up, and the totals are the real ones.
shop-web: 412 items load, 11 ever used here. About 30,600 tokens a request go to things this project never uses.
Every skill, agent, plugin, connector and instruction file, ranked by what it really costs: something small that's sent with every message can cost more than something big that's sent once. It also checks for an AGENTS.md too big for Codex to read in full, instructions one tool sees and the other doesn't, and skills kept where Claude Code can't find them. The card is made on your computer, project names are hidden unless you choose, and loadout never uploads it. For Codex, the split per item is an estimate and is labelled as one.
- Your full setup52,400
- This project's plan21,800
- Difference
- −30,600
Press “Prove it” and loadout runs a few short test chats (six, two tiny messages each), taking turns between your full setup and the plan. It shows the difference only when it's clearly bigger than the normal run-to-run wobble. Those test chats use real tokens, so it only runs when you ask. Everyday chats are measured too, and those numbers are labelled “observed”, because nothing was held the same.
Pick up where you left off
Claude Code in the terminal deletes saved chats after 30 days by default. What you remember or hand off with loadout stays on your computer, so the next chat, in Claude or in Codex, starts where the last one stopped.
- you /loadout:handoff
- claude Saved a handoff note: sessions moved to signed cookies; refresh tokens rotate every 7 days; still open: rate-limit the login route. Branch auth-refactor, 6 files changed. No secrets found.
- next morning, in your terminal
- you loadout continue --in codex
- loadout Here's the note that goes to Codex (412 characters). Send it? y
- codex Picking up the auth refactor: refresh tokens rotate every 7 days. Next open item: rate-limit the login route. Start there?
/loadout:remember
$loadout-remember in Codex
The chat writes a short summary: decisions, open items, file paths. loadout keeps it on your computer for good. Nothing runs in the background.
/loadout:recall
$loadout-recall in Codex
Brings it back in a later chat in the same project, in the same tool. To carry it to the other tool, hand it off.
/loadout:handoff
$loadout-handoff in Codex
Packs where you are into a short note: what's done, what's still open, and facts from git like the branch and the changed files. It checks for secrets before saving anything.
loadout continue
--in claude|codex in your terminal
Opens a fresh chat in Claude or Codex with that note and the project's plan. Before a note goes to the other company's model, you see exactly what will be sent, and nothing goes until you say yes.
Where it works
What each place gets in the first full release. None of it ships yet.
| Where you start a chat | At v1.0 | Details |
|---|---|---|
| Claude Code (terminal) | Full | Audit, a learned plan for every new chat, memory and handoff. |
| Codex (terminal) | Full | Audit, a learned plan for every new chat, memory and handoff. One difference: Codex has no name-only level yet, so skills a project doesn't use stay listed. The total is exact; per-item numbers are estimates. Getting back something that was cut means a quick restart. |
| Claude desktop app (Code tab) | Partial | Audit. From v0.3, turn it on for a project and its chats start with the plan too, with undo, if testing confirms the app picks it up. If not, loadout shows the setting for you to paste. In-chat commands are being tested. |
| Codex app | Not tested | |
| claude.ai and Cowork | Not covered | Outside tools can't reach those chats. |
| Gemini CLI, Cursor, opencode | After v1.0 |
Also after v1.0, if people ask for them: switching accounts per folder, and backing up every chat with search.
What you'll get, in order
First, a short test on real chat history checks that the learning holds up. If it misses too often or cuts too little, v0.2 ships plans you pick yourself and everything else stays the same. Each step ships on its own. No dates until there's something to show.
- v0.1AuditNextWhat every chat loads, what each project has actually used, and what the rest costs. A card you can share, with project names hidden unless you choose. Claude first, Codex right after.
- v0.2Autopilot in the terminalTyping claude or codex in a project starts the chat with that project's plan. One command brings back anything cut, and your full setup is always one flag away.
- v0.3Desktop per-project tuningTurn it on for a project and the Claude desktop app's Code tab starts chats there with the plan too, with undo. Only if testing confirms the app picks it up.
- v0.4Memory and handoffRemember, recall, and hand off to a fresh chat in Claude or Codex. Kept on your computer.
- v0.5Install and deleteSee what something will cost before you add it. Each change asks first and can be undone.
- v1.0PolishWindows, a clean uninstall, onboarding and docs.
After v1.0: accounts, full chat backup and search, if people ask for them.
Questions
Why does a session use so many tokens before I type?
Every new chat sends Claude's own instructions and tools, plus everything you added on top: the list of your skills and agents, connector tools, notes that plugins add, and your instruction files. One public report showed about 56,800 tokens after typing “hi”. Another showed a first request of 72,624 tokens with 247 skills installed, before any work.
How does loadout decide what a chat starts with?
It reads the chat history Claude Code and Codex already keep on your computer and notes what each project (a git repo, or a folder) has used. Anything used there in the last 30 days or the last 10 chats loads in full, and so does anything you start on purpose or pin with “loadout keep”. Skills the project hasn't used are shortened to their name, so Claude still sees them and can call them. Codex has no name-only level yet, so there they stay listed. A plugin or connector is switched off for a chat only if the project hasn't used it in 30 days and has at least 8 chats of history. These are plain written starting rules, and the first tests on real chat history will tune the numbers. It's not a model guessing, and /loadout:why shows the reason for every item.
What if it cuts something I need?
Type /loadout:need and the name. In Claude Code, a skill should come back in the same chat; testing will confirm that before it's promised. In Codex, and for a plugin or connector, loadout gives you one command that restarts the chat with it on. “loadout run --full” starts a chat with your full setup. Some things are never cut automatically: plugins that check or approve actions before they run, and instruction files like CLAUDE.md and AGENTS.md. For a plugin whose use can't be seen in your chat history, loadout shows what it costs and you decide once per project.
Does loadout change my Claude Code or Codex settings?
Not your main ones. Your global Claude Code and Codex settings stay as they are. Only install and delete (v0.5) change them, and only when you click and confirm. The plan applies when a chat starts, through the command that starts it, and only to that chat. loadout doesn't edit your shell startup file either: it prints a short snippet for you to add. One exception, only if you turn it on for a project: the Claude desktop app's Code tab can only be set per project, so from v0.3, if testing confirms the app picks it up, loadout can keep that project's local settings file (.claude/settings.local.json) in line with the plan. It changes only the switches for skills, plugins and connectors, keeps a backup, can undo, leaves alone anything you changed yourself, and won't write the file unless git ignores it. While it's on, every chat in that project uses the plan, terminal chats too, until you turn it off. If testing doesn't confirm it, loadout shows the setting for you to paste.
Claude Code already has /usage and /skill-doctor. Why loadout?
Those are useful, and loadout works alongside them. On Pro, Max, Team and Enterprise plans, /usage breaks recent usage down by skill, subagent, plugin and connector. /skill-doctor shows what each skill costs, flags skills you've never used and says where to turn them off. And when the skill list gets too long, Claude Code already drops the descriptions of the skills you use least, keeping every name. loadout leaves that part alone. It goes after what those leave to you: plugin notes that can be added to every message, connectors, agents, and all of Codex. And it's built to act by itself, per project, in every new terminal chat (and in Desktop projects you turn on), instead of leaving the trimming to you.
Does it read my login or API keys?
No. You sign in with Claude's or Codex's own login. loadout never opens the files or keychain where they keep it, and never stores or passes on your password or login token. It doesn't need an API key.
When do account switching and chat backup come?
After v1.0, and only if people ask for them. Until then, loadout shows you the Claude Code setting that keeps chats longer than 30 days, for you to change; it never changes it for you. Anything you remember or hand off with loadout is kept on your computer either way.
When can I use it?
Not yet. Nothing has shipped. First comes a short test on real chat history to check that the learning holds up, then the audit is the first release. Join the waitlist and you'll get one email when it ships.
Who's building it?
Prashant Dasari, a frontend engineer based in India. More on the about page.
Get the first release
The audit comes first: what every chat loads, and what each project actually uses.
Your email is used for one thing: telling you when loadout ships. No newsletter, no sharing. Ask to be removed any time via the contact page.
loadout is in development. Built by Prashant Dasari.
Back to top