canerganis/akor

Your AI agents keep repeating themselves. AKOR stops you paying for it.
**41% to 43% fewer tokens per verified job on Claude Sonnet 5.5, Opus 5.5 and Fable 5.1, 31% on Haiku 5.5. Faster replies on all four. 91 of 91 replies accepted on the first try.**
Try it in 60 seconds · Results · How it works · FAQ
If AKOR saves you tokens, a star helps other people find it.
The 30-second version
**The problem.** Multi-agent pipelines pay again and again for the same thing: every call resends the rules, the
context and a verbose reply format, and one malformed answer forces a retry.
**The fix.** AKOR packs 64 ready jobs into one call. The goal, each operation and each shared input are written
once, and jobs point at them by index. Reads, tests and status events never touch a model. The model answers one
short record per job, and the host checks the whole reply before it delivers anything.
**The proof.** Live, preregistered tests on billed tokens, with hidden reasoning and failed calls counted, against
a plain JSON format carrying the same 64 jobs in one call. Nothing in the reply is repaired, so every correct job is
really correct.
Who it is for
- You run **several agents or many small model tasks** (reviews, extractions, checks, triage) and the token bill or
the rate limit is what stops you from doing more.
- You use **Claude through claude -p** and want batching, retries, a durable queue and a cost ledger without
writing them yourself.
- You care about **numbers you can audit**: every call is logged with tokens, time and validation.
If you send one prompt at a time and wait for each answer, AKOR will not save you anything; savings start at a few
dozen jobs per call.
Try it in 60 seconds
**With a real model.** The bundled adapter runs each packet through your existing claude -p login. No API key.
--min-batch and --max-wait let small jobs wait briefly for a bigger, cheaper call. Any program that reads one JSON
request on stdin and prints one JSON reply can be an adapter; see USAGE.md.
What AKOR does differently
**It cuts the bill for agent-to-agent traffic by about 40%, and proves it.** 64 jobs travel in one call. The goal,
each operation and each shared input are written once; jobs point at them by index. On Claude Sonnet 5.5 that is
40.7% fewer tokens per verified job than plain JSON, on Opus 5.5 42.8%, on Fable 5.1 40.6%, on Haiku 5.5 30.8%, and all four answer faster.
**Every number is a billed number.** Most token-saving claims count tokens with a tokenizer and stop there. AKOR's
counts are what the provider billed: prompt, CLI overhead, visible answer and hidden reasoning, with failed calls
included. Thresholds, call budgets and data were frozen and hashed before the first call of every test.
**Nothing is repaired.** No quote fixing, no stripping of chatter, no rescue of half-broken lines. A reply is either
exactly right or rejected. On the final wire, 91 of 91 Claude replies were accepted on the first try.
**Small fixes for real failure modes.** A model would sometimes slip on the very first character after it finished
thinking, and one wrong bracket cost a whole batch. The fix is a six-character header line, OZ/1g, that costs 0.2%.
An ambiguous job ("an array holding one pair") produced a flat answer once; now result shapes are stated with a
literal example.
**Deterministic work never reaches a model.** Reads, test requests and status events run on the host. A batch made only
of such work finishes with zero model calls.
**Light.** Standard-library Python, a SQLite queue with resume, a ledger of every call, and no API key: it runs on the
claude -p login you already have.
Results
On a live, pre-registered benchmark where every job is checked by a deterministic oracle, AKOR used
**41% to 43% fewer total tokens per verified job** than plain JSON carrying the same 64 jobs in one call, on Claude Sonnet 5.5,
Opus 5.5 and Fable 5.1 with 64 jobs per call, and 31% on Haiku 5.5, with every reply accepted on the first try and every job