Splits the bill
Bill X-ray: re-reading, cache writes and output, then long chats, cold resumes and model switches.
- Long-chat tax$412
- Cold resumes$96
- Model switches$31
Run a dozen agents side by side and see which one waits for you. Fleet also shows where every token and dollar went.
Free for 3 panes. Pro is $24 at launch, paid once. Apple Silicon, macOS 13 or later.
Running the test suite ✓ 212 files scanned ✓ invoice.py: 4 tests added
Added the retry wrapper and 4 tests. All 31 pass.
Do you want to make this edit to server.js? ❯ 1. Yes 2. Yes, allow all edits 3. No
Updated the README for the new retry flags.
Reading the deploy config ✓ 3 services found
Fixed the login screen layout. 9 tests pass.
Each pane shows WAITING, DONE or working, with its cost and context, read from Claude Code's own hooks. A sound, a Dock count and ⌘J take you to the next one.
Fleet shows the files each agent touched, opens its folder in your editor, and finds old worktrees and caches that can go to the Trash.
Fleet reads your own transcripts on your Mac. Nothing is uploaded.
Bill X-ray: re-reading, cache writes and output, then long chats, cold resumes and model switches.
Cache clock: a countdown on every paused chat, 5 minutes or 1 hour, before the next turn writes it all again.
Tools tab: tokens each skill and plugin costs a turn, its calls in this chat, and one switch for the unused ones.
Context blocks: remove a block, topic or keyword, and a new chat carries the rest word for word. Its quality is not measured.
We set each pass rule before the run, and most ideas failed. We do not claim that Fleet will save you money. The X-ray shows your numbers; this table shows ours.
Each new chat draws on or off for every lever, and Fleet compares both groups on your work, with guardrails.
{"chat":"c42","hc":0,"sm":1,"cap":1}Pinned rules such as "Ask before you push" live in the system prompt, which compaction never rewrites.
• always ask before you pushThe Share tab opens a GitHub issue with totals only. You read it first and press send.
No project names. No chat text.Others found the same trap. In a 2,908-run study of Claude Code, one compression setup cut tool output by 38.4% and raised the billed cost by 6.8%. Prompt caching is the reason.
| Idea | Test | Result | Verdict | What we did |
|---|---|---|---|---|
| Fused search and edit tools | 12 paired tasks with strict checks | About 11% more expensive | Failed | Removed the claim |
| All read-only helpers on Haiku | 14 tasks, blind judge, control arm | 98% cheaper. Score 7.0 vs 8.07 of 10. Worst task lost 4 | Failed | Not a default |
| All read-only helpers on Sonnet | Same 14 tasks | 65% cheaper. Score 6.71 vs 8.07. Worst task lost 5 | Failed | Not a default |
| Haiku for mechanical lookups and log searches only | 14 new tasks, scored by computer | Same answers as Opus on all 14, 97% cheaper. These tasks were easy for every model | Passed, narrowly scoped | Saver's lookup helper uses it. Not a claim about your bill |
| Compact at 100K, no tools | 10 runs on 6 real chats. A worst-case memory quiz | Recall 3.9 vs 6.3 of 8 | Failed | Off by default |
| Compact on resume at 200K | 5 real chats, real next message, control arm | Facts 7.4 vs 8.0 of 8. Next-step plans 1.2 vs 2.0 of 2 | Facts passed, plans failed | On, with the result shown in Settings. You can turn it off |
| Pinned rules: standing rules kept in the system prompt | Paired test, 4 rules checked by script, rule set first. Chat rules vs pinned rules after compaction, plus a no-compaction control | Rules said in the chat were broken 9% of the time after compaction. Pinned: 0%. Control: 6% | No measured benefit | Optional. We do not claim it improves rule-following |
| Compact only after a pause, not at a fixed size | Replay of 30 days of one person's chats. No model calls, quality not measured | Replay: about 23% cheaper for chats and no compaction in the middle of a tool loop. A plain 200K cap: about 18% and 15 mid-loop compactions | Replay only, not proven | Opt-in "Smart compact". Fleet logs each setting change so the effect can be checked on real use |
Small samples, on one person's work, at API list prices. The last row replays past chats: it shows where a rule would have saved money, not what it does to quality.
Fleet does not resell tokens, so you pay us once and your vendor as usual.
$0
$24launch price, $39 after
Local tax may apply at checkout. Payments go through Dodo Payments.
Fleet runs the official claude and codex commands on your Mac. It does not modify them or read your login tokens.
READYFleet 1.0.1 · notarized by Apple
Fleet is signed, notarized by Apple and passes 278 automated checks. It needs macOS 13 or later on Apple Silicon. Open the DMG and drag Fleet to Applications.
Download the DMG, 5.3 MBIt downloads the app, checks its SHA-256 and puts it in Applications. Read the script first.
curl -fsSL https://fleetapp.in/install.sh | shOr take the DMG from the latest release.
Yes. It runs Claude Code in your own account, the same as a terminal. You sign in through Claude Code, not through Fleet.
If those work for you, keep them. Fleet is for when you run enough agents that finding the one waiting for you has become the job.
We do not promise that. See the test table. The X-ray shows where your usage goes, and what you change is up to you.
No. Fleet is an independent app and is not affiliated with or endorsed by either company. Claude Code and Codex belong to their owners.