1
Why agents burn tokens
To change one function, an agent usually opens the whole file it lives in, then the files it imports, then opens some of them again a few steps later. Every line it reads is billed, every time. On a real task in Sweepspace's own code, the files an agent would open added up to 35,777 tokens, most of them lines it never needed.
2
What the map does
Sweepspace indexes your project into files, imports and symbols, and serves that map to your agents as MCP tools. Instead of opening a file, the agent asks a precise question and gets a precise answer. Measured on Sweepspace's own code:
| Tool | What it returns | Tokens | Whole file |
|---|---|---|---|
overview | The whole project in one answer | 445 tokens for 64 files | – |
search | A function or file by name, with its line | 34 | – |
outline | A file's functions and types, no bodies | 1,144 | 10,656 |
read_symbol | One function, not the file around it | 640 | 10,656 |
read_lines | Just the lines it needs | 532 | 10,684 |
context | A file with only the parts of its neighbours it uses | 128 | 3,543 |
check | Type-check, only the errors | 21 | 3,319 |
changes | What changed, by file and function | 391 | 5,694 |
The answers also say where their knowledge ends ("not indexed: what macros expand to"), so an agent knows when to look further instead of assuming it saw everything.
3
How it was measured
The task: show merge conflicts on the Board, in Sweepspace's own codebase (64 files), in twelve steps from getting an overview to type-checking the result. The baseline opens each file it needs once, in full, and reads the compiler's full output. With the map, the same steps took 3,833 tokens instead of 35,777: 9.3 times fewer. Reviewing what changed took 14.6 times fewer, and type-checking 158 times fewer.
Your projects will differ, so measure your own: in Sweepspace, open Savings and press Measure. The real tools run on your project and show your ratio.
4
Set it up in two minutes
- Download Sweepspace for Windows (free, no account) and open your project folder.
- Start Claude Code, Codex, Gemini CLI or OpenCode from Sweepspace and connect the map to it in one click (the Pro beta connects every installed agent by itself).
- Work as usual. The Savings screen counts every token the map saved.
5
More ways to spend less
- Savings switch. Long command output from Claude Code is squeezed to what matters, only for commands your own permission rules already allow.
- Free models. OpenCode runs free models with no account. Run it in a terminal next to Claude Code for the easy jobs.
- Hand off easy work. In the Pro beta, Claude Code can hand easy and medium tasks to OpenCode's free models, each on its own branch, and merge the result.
- Pick the model per step. In the Pro beta's crew, Ledger gives the top model only to the builder: in tests, the same tokens as one top-model agent, about 17% cheaper.
6
Questions
Does Sweepspace work with Claude Code?
Yes. Sweepspace connects its map to Claude Code in one click, and to Codex, Gemini CLI, OpenCode and Antigravity too. Claude Code then calls tools like outline and read_symbol instead of opening whole files.
Is the 9.3x number real?
It was measured on one real task in Sweepspace's own code (64 files) on 30 September 2026: 35,777 tokens when each file is opened in full, 3,833 with the map. Your projects will differ, so the app measures your own: open Savings and press Measure.
Does my code leave my computer?
No. The map is built and served on your PC. Your agents keep talking to their own providers, as they already do.