Shape Rotator Cohort · Summer 2026 · Draft — awaiting founder sign-off

BITROUTER

The self-improving LLM router. Set a goal — accuracy, cost, latency — and it routes every call of the agent loop to hold it, getting cheaper with every run.

What they're building, who they are, and where each claim comes from — drawn only from the four source tiers below. Program sessions stay internal unless the team approves an excerpt.

POSTER / PITCH stated by the team WEB publicly checkable REGISTRY cohort records AI synthesis — challenge it Pending approval needs founder sign-off
Product
BitRouterself-improving LLM router
Founders
Kelsen Liu · Archer YangNYC / HK
Benchmark
−33% cost per runterminal-bench-2.1 · their measurement
Open source
Rust · Apache-2.0211 stars · checked 07-22
See it
bitrouter.ailive · checked 07-22
01

The problem

"Tokenmaxxing — the biggest model on every call — buys accuracy by sacrificing cost & latency." POSTER

To hit accuracy, tokenmaxxing spends its way into one corner of the accuracy–cost–latency triangle — and the pull compounds from a single call to a full agent run. "One lever — more tokens — can't hold three constraints at once. A minor trade per call becomes a structural miss per run." DECK

Static routers

Pick one model up front, never adapts. POSTER

Per-call auto-switch

Switches per call, but never lets you set the cost / latency / accuracy balance for the whole workflow. POSTER

Split stack

AI gateway and AI observability are separate tools — pay for both and still hand-build the loop between them. POSTER

02

What is it

"Drop BitRouter into any loop — your models, your tools, your harness — then point it at your goal and it optimizes every call toward it." DECK

Built for two people on the poster: Will, a loop engineer who "lives in agentic coding loops — Claude Code, Cursor, Codex — all day," and Acme, a Series-A AI startup "running production agentic loops at scale." The docs put it plainly: an open-source, adaptive LLM gateway that runs as a single local binary, so "every step of every loop stops billing at frontier prices by default." POSTER WEB

03

What it does

Pick a point in the accuracy–cost–latency triangle; BitRouter routes every call to hold it there.

Goal-point routing

"You choose the balance… You set the point — BitRouter holds it, every run." DECK

Bring your own everything

Your harness (ACP, subagents), 100+ frontier and open models, your MCP tools and skills — "one agent loop · any topology." DECK

Routing as code

A declarative routing-policy.yaml — "version it, diff it, ship it. The loop reads it, then rewrites it." DECK

Reliability

Mid-run rerouting and automatic fallbacks keep "long loops alive through outages." WEB

Observability

Full tracing with cost attribution per run, exported over OTLP — in the path, nothing to bolt on. WEB DECK

Guardrails

Guardrails and rate limits "enforced once, at the router, for every agent." WEB

04

How it works

"Not a static router that decides once and rots — BitRouter runs a closed loop on every call, so it gets cheaper and sharper each lap with no tuning by you." DECK

Policy in

optimize: costmin_accuracy 0.95 max_latency 8 sbudget $2.00/run
The BitRouter loop
Act — route each call to the model that fits Observe — trace cost, latency, outcome per run Evaluate — score complexity vs the threshold Learn — rewrite threshold & model mix
CHEAPER EVERY RUNthe rewritten policy compounds
ANY HARNESSClaude Code · opencode · codex · goose

Routine work goes to open models; the hard calls escalate to frontier — the deck's example policy defaults to open/qwen3-coder, escalates past complexity 0.7 to frontier/claude-sonnet, and halts on spiral. DECK

Built with: Rust, Apache-2.0 (stated open source — the public repo checks out) · a single local binary · OTLP tracing · curated open-weight models: Qwen3, Llama, DeepSeek, Kimi K2, GPT-OSS. The poster's ecosystem tiles: DeepSeek, Qwen, Kimi, MiniMax, StepFun, Xiaomi MiMo, Z.ai, Tencent. POSTER DECK WEB

05

The numbers

The team's own benchmark and figures, as printed on their poster and deck.

−33%
Cost per run · terminal-bench-2.1 · quality held
77 → 76
Solve rate, frontier solo vs routed — their mock-pitch figures
$0.026
Poster's example workflow total · p50 88 ms
0%
Token markup · open-source core, self-host free

The poster goes further: "SOTA on router benchmarks — beats closed-source routers like Droid Router & NotDiamond." Live checks: bitrouter/bitrouter is public — Rust, Apache-2.0, 211 stars, pushed 2026-07-22 — and a BitRouterAI/benchmarks dataset is public on Hugging Face (updated mid-July). The benchmark results themselves are the team's measurements. POSTER WEB Pending approval

Two things to reconcile before sign-off. The deck's traction chart is labeled "measured, not modeled" yet watermarked "placeholder data — for layout review" — its per-model bars are not quoted on this page. And the same $0.026 example workflow reads "−33% cost" on the poster but "−80% vs. all-frontier baseline" on bitrouter.ai (fetched 07-22). Conflict — confirm Conflict — confirm
06

Who's building it

Two co-founders, NYC / HK. Deck framing: "We are the routing layer we're building" — Chinese frontier labs as supply and origin, the global OSS community as demand and roots. DECK REGISTRY

Kelsen Liu

Co-founder & CEO

NYU — CS, Media & Business. Self-taught full-stack engineer. 3× consecutive ETHGlobal winner. — their deck's team slide.

x.com/kelrouter github.com/SPIKESPIGEL404 Confirm
POSTERDECKREGISTRY

Archer Yang

Co-founder & CTO

Zhejiang University — Aerospace. China NOI medalist. Rust China Conf speaker. Full-stack — ML, blockchain, embedded. — their deck's team slide.

x.com/arcrouter
POSTERDECK
07

Traction

"From self-improving LLM router to everything" — the deck's timeline, checked against what's live.

Gateway & observability

Their milestones: 3B+ monthly token usage, 2,000+ downloads, 200+ GitHub stars. Live: 211 stars, 13 forks on bitrouter/bitrouter, pushed 2026-07-22. DECK WEB Usage figures pending

The benchmark

"SOTA performance on Terminal-bench-2.1," routing traces open-sourced. The Hugging Face dataset BitRouterAI/benchmarks is live (updated mid-July; checked 07-22). DECK WEB

The org

19 public repos: harness integrations (OpenClaw, Hermes Agent, Pi Coding Agent), an elizaOS provider plugin, a Homebrew tap, and an open eval database of models and harnesses. WEB

The business: "Open source is the distribution — cloud is the business." Adopt at 0% markup (self-host free), a $20/mo retail-dev bundle of SOTA open-weight models, managed cloud and enterprise above it — "you pay list; we buy at a volume discount — the spread is our margin." Next on the timeline: an optimizer learned from every run's traces via distributed RL, then a full RL stack for personalized workflows. DECK

Cohort records also list x402-kit — an HTTP-native micropayment SDK for agents — and a cohort workshop on agentic payment systems among the team's offerings. REGISTRY
08

In their own words

How the team tells it — from their Mock Demo Day pitch, 2026-07-17, Kelsen Liu presenting.

"Everyone else routes the call. Only BitRouter turns routing, observability, evals and guardrails into one self-improving loop." The closing line of the deck's competition slide DECK

"In any agentic workflow in production it's cost, speed, or quality — how do you balance it is the problem we are trying to solve."

PITCHPending approval

"We route each call to the model that fits, then we observe it, then we evaluate it — then our router learns the whole workflow and tunes the policy to optimize it."

PITCHPending approval

"We benchmarked our router paired with GPT-5.5 on terminal bench 2.1 — 33% less cost without any performance loss. The control, purely GPT-5.5, is 77; ours is 76 at a third less cost."

PITCHPending approval

"You can run any agents you use every day wrapped inside the TUI — at the bottom you see what a session costs — and let the optimizer agent observe and optimize the whole workflow for you."

PITCHPending approval
09

Try it & links

Everything below was live-checked on 2026-07-22 unless marked otherwise.

SurfaceWhereSource
Websitebitrouter.ai — "A self-improving LLM router that optimizes agentic workflows with every run."WEB Live · 07-22
Docsbitrouter.ai/docs — quickstart, models, routable primitives; core "open-source under Apache 2.0 and self-hostable for free"WEB Live · 07-22
GitHubgithub.com/bitrouter/bitrouter — Rust · Apache-2.0 · 211 stars · pushed 07-22WEB Live · 07-22
Benchmarkshuggingface.co/datasets/BitRouterAI/benchmarks — the deck's open-sourced tracesWEB Live · 07-22
Xx.com/BitRouterAI — on the poster and linked from the site; direct fetch blocked 07-22POSTER REGISTRY Confirm live
Communitydiscord.gg/G3zVrZDa5C · founder call: cal.com/kelsenliu · linkedin.com/company/bitrouterai (registry; not live-checked)DECK REGISTRY
Contactkelsenliu@bitrouter.ai · archeryang@bitrouter.ai (poster) · contact@bitrouter.ai (site)POSTER WEB Confirm preferred
Demo link — empty slot
The Demo Day links sheet has no Bitrouter row and the registry's demo field is blank (checked 07-22). Add the demo URL when it exists.
Unrelated namesake: bitrouter.com and the @bitrouter handle belong to a broadcast-TV company founded in 2002 — not this team. REGISTRY
10

Provenance ledger

Every load-bearing claim and where it comes from; decisions live in §11.

ClaimComes fromTierStatus
One-liner, tokenmaxxing framing, the closed loop, primitivesCohort poster + demo-day deck, July 2026POSTER DECKTeam's own artifacts
Founders, roles, bios, NYC / HKPoster team block + deck team slide; registry corroborates (2 members, NYC / HK)POSTER REGISTRYCorroborated
−33% cost per run on terminal-bench-2.1, quality held; Droid Router / NotDiamond comparisonPoster proof panel; deck; repeated in the mock pitch — team's own measurements, named third partiesPOSTERPending approval
Deck traction chart (per-model bars)Deck p9 — labeled "measured, not modeled" yet watermarked "placeholder data — for layout review"; bars not quoted hereDECKConflict — confirm
$0.026 example: −33% (poster) vs −80% vs all-frontier (site)Poster policy.yaml panel vs bitrouter.ai, fetched 07-22 — same example, different baselinePOSTER WEBConflict — confirm
Repo facts (Rust, Apache-2.0, 211 stars, pushed 07-22, 19 org repos)api.github.com, fetched 2026-07-22 — corroborates the deck's "200+ stars"WEBVerified live
3B+ monthly tokens, 2,000+ downloadsDeck traction timeline — no public counter found to checkDECKPending approval
Benchmarks dataset publichuggingface.co/BitRouterAI — dataset "benchmarks," updated mid-July, fetched 07-22; contents not auditedWEBVerified live
X handle @BitRouterAIPoster, registry, and the site's footer; direct fetch returned an access wall 07-22POSTER REGISTRYConfirm live
Earlier registry framing: "P2P LLM router," x402-kit, agentic-payments workshopCohort registry — records note the team repositioned around reliability + cost for coding agents; current poster/deck say "self-improving router"REGISTRYConfirm framing
Pitch excerpts (§08)Mock Demo Day, 2026-07-17; lightly punctuated, two ASR garbles decoded (noted in tooltips)PITCHPending approval
11

Founder review

Five things before this page goes public.

The deck's traction chart — which numbers are real?It says "measured, not modeled" but carries a "placeholder data — for layout review" watermark. This page quotes only the headline −33%; confirm or supply the real per-model figures.
Open
−33% or −80% on the $0.026 example?Your poster prints "−33% cost"; bitrouter.ai shows the same example at "−80% vs. all-frontier baseline." Tell us which baseline this page should quote.
Open
Your team, listed right?Bios are from the deck's team slide; handles @kelrouter / @arcrouter from the poster, Kelsen's GitHub from the registry. Preferred contact: kelsenliu@, archeryang@, or contact@?
Open
The earlier P2P / payments story — keep any of it?Cohort records still list "P2P LLM router," x402-kit, and the agentic-payments workshop. Say the word and it stays; otherwise the self-improving router is the whole public story.
Open
Anything you'd cut or fix?Likely spots are marked "Pending approval": the benchmark figures, the usage numbers, the pitch excerpts, the contact line.
Open