Live Beta — 30 days free

Stateful reasoning that survives interruption

LARS is the first LLM runtime that preserves reasoning state across user interrupts. Powered by the Dabab token-saving layer — 50-75% less cost.

lars-cli · POST /api/v1/chat200 OK

POST /api/v1/chat
{
  "prompt": "Plan a SaaS launch in Egypt",
  "active_overrides": ["focus on Cairo only"],
  "llm_tier": "mini"
}
→ State preserved. 64% tokens saved.

Live platform metrics

Real-time data from the LARS Live Beta. Updates every 10 seconds.Live

+12% wk

1,247

Total Calls

64.3%

337,749

Tokens Saved

L1 semantic

7.1%

Cache Hit Rate

-18% wk

982ms

Avg Latency

Architecture

How LARS works

A 5-stage pipeline that preserves reasoning state and saves tokens at every layer.

User Request

Prompt + active_overrides

FastAPI Gateway

Auth, rate limit, route

Dabab Layer

Cache + compress + predict

LARS Core

State merge + intent handler

Production LLM

gpt-4o-mini / standard / frontier

L1 Semantic Cache

0 tokens on repeat. Exact + fuzzy match (Jaccard 4-gram, threshold 0.65).

Semantic Compression

~80% smaller context. Local Llama-3-8B preserves entities + numbers.

Intent Prediction

Pre-warm likely next. Local Llama-3 predicts next-intent, caches ahead.

Typed Interrupts

9 typed interrupts — not free-text chat

Every user message is classified into one of 9 typed intents. Each has a deterministic merge handler — no hallucinated state changes.

LiveSCOPE_NARROW

"focus on Cairo only"

Rewrites broad-scope references

PlannedSCOPE_EXPAND

"also include the Gulf"

Inserts new pending step

LiveCORRECTION

"actually use blue"

Modifies most recent step

LiveREPLACE

"use Twitter instead of Facebook"

Swaps token everywhere

LiveADD

"also include TikTok"

Appends pending step

LiveREMOVE

"drop the influencer budget"

Drops matching step

PlannedREPRIORITIZE

"do budget first"

Re-ranks pending steps

LiveCLARIFY

"what do you mean by young?"

No-op + log

LiveABORT

"stop, restart"

Clears state

The only method that preserves + adapts

12 tasks × 4 methods. LARS is the unique Pareto-optimal point.

Method	RPR ↑	Cost ↓	Used interrupt?	Win?
`no_interrupt`	1.000	62.4	✗	✗
`restart_from_scratch`	0.000	67.7	✓	✗
`langgraph_checkpoint`	0.000	67.7	✓	✗
`lars`	1.000	62.4	✓	✓

Reproduce with: python examples/run_benchmark.py

Pricing

Simple pricing

Start free. Upgrade when you scale.

Free

For testing the Live Beta

$0/month

1,000 calls/month

All 9 LARS intent types
3-layer merge pipeline
L1 semantic cache
gpt-4o-mini (mini tier)
Community Discord support
50-75% token savings

Ready to try LARS?

Get 1,000 free calls during our 30-day Live Beta. No credit card required.

Running on a 16GB server in Cairo. Powered by Llama-3-8B + gpt-4o-mini.