Atlias
Atlo, the Atlias mascot
Atlias

Loading Atlias

Next Gen coding harness agent

Same model. A fraction of the tokens.

Atlias is a coding agent harness that checks every claim of done, remembers your project, and spends up to twenty times fewer tokens than plain Claude Code on the same tasks.

npm install -g github:ridelink0/atlias
Get Atlias

02 Proof

Benchmarks

Measured on one local model, qwen3.5:9b, on the same tasks with the same grader: Atlias's own agent, and Claude Code with sub-atlias, each against plain Claude Code.

Tokens per solved task

Tokens per solved task: Atlias and sub-atlias against plain Claude Code2 record(s) with both token totals drawn; 9 without totals listed below the chart. Screen 1 of 9, October 17 qwen3.5:9b, 32k window, effort default Atlias 10,195 tokens per solved task Claude Code (plain) 246,712 tokens per solved task Sample: 120 runs. Claude Code with sub-atlias, ten-task smoke, October 16 qwen3.5:9b, 32k window, effort default Claude Code + sub-atlias 31,268 tokens per solved task Claude Code (plain) 82,389 tokens per solved task Sample: 10 runs.
Text table
RecordArmSolvedTotal tokensTokens per solved task
Screen 1 of 9, October 17Atlias86 of 120876,81010,195
Screen 1 of 9, October 17Claude Code (plain)80 of 12019,736,971246,712
Claude Code with sub-atlias, ten-task smoke, October 16Claude Code + sub-atlias8 of 10250,14331,268
Claude Code with sub-atlias, ten-task smoke, October 16Claude Code (plain)8 of 10659,11282,389
Screen 2 of 9, October 17Atlias83 of 120not measurednot measured
Screen 2 of 9, October 17Claude Code (plain)80 of 120not measurednot measured
Screen 3 of 9, October 17Atlias79 of 120not measurednot measured
Screen 3 of 9, October 17Claude Code (plain)80 of 120not measurednot measured
Screen 4 of 9, October 17Atlias82 of 120not measurednot measured
Screen 4 of 9, October 17Claude Code (plain)80 of 120not measurednot measured
Screen 5 of 9, October 17Atlias89 of 120not measurednot measured
Screen 5 of 9, October 17Claude Code (plain)80 of 120not measurednot measured
Screen 6 of 9, October 17Atlias85 of 120not measurednot measured
Screen 6 of 9, October 17Claude Code (plain)80 of 120not measurednot measured
Screen 7 of 9, October 17Atlias85 of 120not measurednot measured
Screen 7 of 9, October 17Claude Code (plain)80 of 120not measurednot measured
Screen 8 of 9, October 17Atlias83 of 120not measurednot measured
Screen 8 of 9, October 17Claude Code (plain)80 of 120not measurednot measured
Screen 9 of 9, October 17Atlias82 of 120not measurednot measured
Screen 9 of 9, October 17Claude Code (plain)80 of 120not measurednot measured
Ten-task smoke, October 16Atlias9 of 10not measurednot measured
Ten-task smoke, October 16Claude Code (plain)8 of 10not measurednot measured

Nine screens, one dot each

Total-token reduction on each of the nine screensEach dot is one 60-task screen at its ratio of total tokens, plain Claude Code over Atlias. The rule marks 20x. The pooled ratio is 19.52x. 0x5x10x15x20x25x 20x pooled 19.52x Screen 1 of 9, October 17: 22.51x22.51x Screen 2 of 9, October 17: 22.36x22.36x Screen 3 of 9, October 17: 14.02x14.02x Screen 4 of 9, October 17: 15.48x15.48x Screen 5 of 9, October 17: 19.66x19.66x Screen 6 of 9, October 17: 19.63x19.63x Screen 7 of 9, October 17: 23.42x23.42x Screen 8 of 9, October 17: 20.95x20.95x Screen 9 of 9, October 17: 22.91x22.91x
Text table
ScreenRatioAtlias solvedPlain Claude Code solved
Screen 1 of 9, October 1722.51x86 of 12080 of 120
Screen 2 of 9, October 1722.36x83 of 12080 of 120
Screen 3 of 9, October 1714.02x79 of 12080 of 120
Screen 4 of 9, October 1715.48x82 of 12080 of 120
Screen 5 of 9, October 1719.66x89 of 12080 of 120
Screen 6 of 9, October 1719.63x85 of 12080 of 120
Screen 7 of 9, October 1723.42x85 of 12080 of 120
Screen 8 of 9, October 1720.95x83 of 12080 of 120
Screen 9 of 9, October 1722.91x82 of 12080 of 120

Pooled ratio 19.52x: a ratio of pooled token totals across the nine screens, not a mean of the screen ratios.

03 How it works

Four steps, every task.

Atlo, the Atlias mascot

  1. Step 1

    It reads only what it needs.

    Its first prompt is 729 tokens before any work. Plain Claude Code starts at about 19,300 to 24,100 tokens on the same tasks.

    > prompt: 729 tokens before any work
> reads: the one file the task names
> ready
    Illustration, not a recording.

  2. Step 2

    It checks before it says done.

    After an edit, Atlias runs the project's check. A reply that says done with no check since the last edit is held back.

    > edit: parser.js
> check: npm test
> 14 passed, 0 failed
> done, after a check that ran past the edit
    Illustration, not a recording.

  3. Step 3

    It remembers across sessions and harnesses.

    Project memory, handoff notes and a consolidation pass carry work from one session to the next.

    > memory: 3 notes loaded
> handoff: resumed from the last session
> consolidated: 2 duplicate notes merged
    Illustration, not a recording.

  4. Step 4

    Your plan or your keys.

    Use ChatGPT, Gemini, an Anthropic API key, Claude Code, or a local model through Ollama. You choose once, on the first screen.

    > route: local model via ollama
> no account needed for this route
    Illustration, not a recording.

05 Works with

Your host, your model.

  • Claude Code
  • Codex
  • ChatGPT
  • Gemini
  • Anthropic API
  • Ollama