Tools
An agent can do almost anything with a shell, and that's the trouble. A find in
a real project returns half a megabyte, a sed that matches nothing reports
success, and the model pays for all of it in tokens and turns.
So every tool in Luxe has to beat the command it replaces: by sending less, by getting it right the first time, or by doing something the shell can't.
Measured in a checkout of Django at 04ac9b45a3, the codebase
the harness eval runs in, with its virtualenv in place. The shell row is the command's
raw output; tightest cut is the least any other agent sends its model of it,
which is Codex's 10 KB, head and tail kept. Pi and OpenCode keep up to 50 KB and Claude
Code 30,000 characters. A token is about four bytes.
Those are single calls, and across a whole task the bars shrink, because most calls are small to begin with. What a whole task costs is below.
Luxe, Claude Code, Codex, OpenCode and Pi each ran the first twelve Django tasks of SWE-bench Verified Mini, with the same model, the same prompt and the same keys, every call metered at the network and priced from one list. Each figure is what the other agent spent on the same task, as a multiple of what Luxe spent: above 1× costs more than Luxe.
| Model | Claude Code or Codex | OpenCode | Pi |
|---|---|---|---|
claude-haiku-5-5 | 1.71× | 1.16× | 0.60× |
claude-sonnet-5-5 | 1.98× | 1.15× | 0.44× |
claude-opus-5-5 | 1.76× | 1.14× | 0.48× |
gpt-6-luna | 1.23× | 0.97× | 0.45× |
gpt-6.1-sol | 1.73× | 1.05× | 0.44× |
Every run is charged as a cold start, the way a session begins after a break. Run side
by side on a warm cache, Claude Code and OpenCode cost up to 2.6× what Luxe does on
Claude models. Pass rates overlapped on every model. Pi is the cheapest by design: four
tools and a short prompt, with web search, MCP and the rest left to extensions. Graded
on one machine rather than in SWE-bench's own containers, so these are not leaderboard
scores. Measured at 5b58b8b0.
instead of cat, head, sed -n
Several files in one call, with numbered lines. An image comes back as a picture.
instead of sed -i
Changes exact text, several edits in one call, all or none. It won't touch a file the agent hasn't read, or one that changed since.
instead of cat > file
Creates or replaces a whole file, and shows what it wrote.
instead of find
Skips whatever .gitignore skips, so build output and dependencies stay out. Sorted by path, capped at 200, with a count of the rest.
$ find . -name '*.py' 27k tokens glob 1.7k tokens
instead of grep -rn, rg
Capped per file and overall, with long lines trimmed. It can return just the files that match or just the counts, and says so when there's more.
$ rg -n 'def delete' 1.1k tokens grep 1.1k tokens
instead of grep on code
Search and replace by code structure. A rewrite shows a preview before it writes.
$X.unwrap() finds the calls and nothing else.
instead of sh
Long output comes back as the lines that matter, with every gap marked and the untouched text saved to a file, each stream whole up to 8 MB. A command that runs out of time returns what it printed so far.
$ git log --stat -n 30 3.1k tokens bash 810 tokens
instead of cargo test, npm test, pytest
One line when everything passes. When something fails, each failure with its file:line, and which ones your change broke. Tests run the project's way, through its own .venv or package manager.
$ cd tests && ../.venv/bin/python runtests.py --settings=test_sqlite --parallel 1 delete 90 tokens test 60 tokens
instead of git, gh
The last ten commits, with a note when there are more. Diffs are capped and say how to narrow them. Pull requests, issues and CI checks come through gh.
$ git log 2052k tokens git 240 tokens
instead of cat
One function, struct or method by name, instead of the file it lives in.
$ cat django/db/models/deletion.py 3.8k tokens read_symbol 420 tokens
instead of cat
A file's outline: each symbol and the lines it spans, so the agent reads only what it needs.
$ cat django/db/models/query.py 20k tokens module_report 490 tokens
instead of rg 'fn name'
Where something is defined, from its name or part of it.
instead of nothing
What else a change touches, before the agent makes it.
instead of cargo check, tsc
Errors and warnings for the files you name. They also come back on their own after every edit.
instead of grep for a name
Definitions, references, types and renames, from the language server.
instead of print statements
Drives a real debugger, with breakpoints, stepping and variables, for Rust, C, C++, Python and Go.
instead of curl
Pages come back as markdown, not HTML. A GitHub repository is cloned into the project, with its links flattened and no language server started in it, so the agent can search the real files.
$ curl -sL https://docs.rs/tokio/latest/tokio/ 14k tokens web_fetch 5.4k tokens
instead of curl a search page
Searches the web through the provider you have a key for, or a free one.
instead of chrome --headless --screenshot
Captures a page, so the agent can check the UI it just built.
A checklist that survives compaction, shown in the dock's PLAN.
Stops to ask you, with choices, instead of guessing.
Loads a skill's instructions when the task calls for them.
Your MCP servers behind one tool, every call checked against your permissions.
Every call has one, and ctrl+esc stops the whole batch.
A call with an argument the tool doesn't know still runs, and says which it ignored.
/tools shows what the agent can use, and any tool but the core four can be
turned off. See the guide.