PLAYGROUND

Test the compression
before installing.

Paste a log, JSON or diff. The real cl100k tokenizer measures input, output and savings on the fly.

Ensaio

log

Tokenizer

cl100k

0→0

0processed

0saved

—total margin

0capsules

0tests

Entrada

0tokens

// nuxs-capsule · compress pipeline
const compress = async (input: string, kind: CapsuleKind) => {
  const tokens = encode(input, 'cl100k_base')
  const cached = await readDedup.lookup(sha256(input))
  if (cached) return cached.capsule
  const groups = cluster(tokens, { method: capsule.kind })
  const capsule = pack(groups, { ratio: configured, kind })
  emit('telemetry', { in: tokens.length, out: capsule.size })
  return capsule
}

interface Capsule {
  type: 'diff' | 'log' | 'codebase' | 'api' | 'prompt'
  input_tokens: number
  output_tokens: number
  ratio: number
  fingerprint: string
}

// [INFO] capsule#1842 · diff · 23,141 → 6,553 tokens · 71.7% saved · 14ms
// [INFO] capsule#1843 · log  · 19,759 → 156    tokens · 99.2% saved · 8ms
// [INFO] capsule#1844 · api  · 17,390 →  45    tokens · 99.7% saved · 6ms

await proxy.forward(req, { capsule: 'auto', tokenizer: 'cl100k_base' })

// nuxs-capsule · compress pipeline
const compress = async (input: string, kind: CapsuleKind) => {
  const tokens = encode(input, 'cl100k_base')
  const cached = await readDedup.lookup(sha256(input))
  if (cached) return cached.capsule
  const groups = cluster(tokens, { method: capsule.kind })
  const capsule = pack(groups, { ratio: configured, kind })
  emit('telemetry', { in: tokens.length, out: capsule.size })
  return capsule
}

interface Capsule {
  type: 'diff' | 'log' | 'codebase' | 'api' | 'prompt'
  input_tokens: number
  output_tokens: number
  ratio: number
  fingerprint: string
}

// [INFO] capsule#1842 · diff · 23,141 → 6,553 tokens · 71.7% saved · 14ms
// [INFO] capsule#1843 · log  · 19,759 → 156    tokens · 99.2% saved · 8ms
// [INFO] capsule#1844 · api  · 17,390 →  45    tokens · 99.7% saved · 6ms

await proxy.forward(req, { capsule: 'auto', tokenizer: 'cl100k_base' })

// proxy.nuxs.ai · multi-provider routing
const upstream = detectUpstream(req.body.model, env)
switch (upstream) {
  case 'anthropic': return forward(req, ANTHROPIC_KEY)
  case 'openai':    return forward(req, OPENAI_KEY)
  case 'google':    return forward(req, GOOGLE_KEY)
  case 'groq':      return forward(req, GROQ_KEY)
}

// diff --git a/server.ts b/server.ts
// + connectionTimeoutMillis: 5000,
// + statement_timeout: 3000,
// + query_timeout: 3000,
// - // no timeout — pool would hang

const record = await pool.query(
  'INSERT INTO compression_events (type, in_tok, out_tok, saved) VALUES ($1, $2, $3, $4)',
  [event.type, event.in, event.out, event.in - event.out]
)

// [METRIC] processed=180,322,482 saved=157,693,455 margin=87.45%
// [METRIC] runs=9569 capsules=17 tokenizer=cl100k_base
// [METRIC] dedup=on tokenizer=cl100k_base
return { ok: true, capsule_id: record.rows[0].id, ratio: capsule.ratio }

// proxy.nuxs.ai · multi-provider routing
const upstream = detectUpstream(req.body.model, env)
switch (upstream) {
  case 'anthropic': return forward(req, ANTHROPIC_KEY)
  case 'openai':    return forward(req, OPENAI_KEY)
  case 'google':    return forward(req, GOOGLE_KEY)
  case 'groq':      return forward(req, GROQ_KEY)
}

// diff --git a/server.ts b/server.ts
// + connectionTimeoutMillis: 5000,
// + statement_timeout: 3000,
// + query_timeout: 3000,
// - // no timeout — pool would hang

const record = await pool.query(
  'INSERT INTO compression_events (type, in_tok, out_tok, saved) VALUES ($1, $2, $3, $4)',
  [event.type, event.in, event.out, event.in - event.out]
)

// [METRIC] processed=180,322,482 saved=157,693,455 margin=87.45%
// [METRIC] runs=9569 capsules=17 tokenizer=cl100k_base
// [METRIC] dedup=on tokenizer=cl100k_base
return { ok: true, capsule_id: record.rows[0].id, ratio: capsule.ratio }

📂

Drop a file here
or click to pick a Nuxs sample

—

compression

Saída compactada

0tokens

paste something on the left

Samples to try

The samples sit below the test zone so they don’t steal the fold.

algorithmic · runs here

The playground is a demo. The free plan is proof.

50 million lifetime tokens, all 20 types, local and no credit card. Install, run on your code, check the dashboard on Friday.

Install the free plannpm install -g nuxs-capsule

How the test works

1 · Pick the type

Log, JSON, schema, session, PR… each type has its own compressor. Click a chip above.

2 · Pick the size

Small / Medium / Large — or paste/upload your own. The BIGGER and more repetitive the input, the bigger the win.

3 · Real compression

The cl100k tokenizer measures input and output on the fly. Algorithmic ones run free here in the browser; LLM-based ones run on the server up to the public limit.

4 · Fully audited

Tokens are real (cl100k), the cost of each test is shown on screen, and the public budget is transparent. No inflated numbers.

This is just to see it work. For a real run on your code, install the free plan — 50M lifetime tokens, local, no credit card.

SCOREBOARD BY CAPSULE17 testable capsules · every test you run above adds to its type’s score

CAPSULEWHAT IT COMPRESSESTOKENS PROCESSEDSAVEDMARGINBEST ×

11 algorithmicclient-side · zero LLM cost · free plan

LGlogNginx/Node access logs — repeated templates, varying timestamps.————

APIapiHTTP request logs (same format the benchmark measures — 99.7%).————

NWnetworkHAR — slowest requests, errors, hosts. Tolerant to annotated exports.————

DFdiffgit diff — signal without the noise.————

TStestMulti-runner (pytest/jest/vitest/playwright) — pass/fail per runner + root causes.————

BDbuildCI pipelines — stage durations, cache hit rate, bundle budgets, canary.————

CBcodebaseFile map + repo exports.————

APIapispecOpenAPI spec → navigable summary.————

DBschemapg_dump --schema-only → essence.————

PRpromptSemantic dedup of repeated prompts.————

IMimageBinary image (base64) → compact reference via magic bytes. Zero LLM.————

6 LLMserver · real provider · paid plans

SQLsqlSlow query log — patterns by shape, pathological queries, deadlocks.————

STstackStack traces — causal chain + raw context lines around each frame.————

EVeventsMixpanel/PostHog — funnel, anomalies w/ time window + tenant/campaign dims.————

PDFpdfStructured PDF → in-context summary.————

RAGragRerank chunks by relevance.————

THthreadsPR/Slack threads → decisions by channel + incidents with root cause.————

Aggregate total:0 processed0 saved— global margin0 capsules0 tests

Multimodal capsules

Video, image and meeting become dense text ready for LLM context. Testable here soon — for now, typical product expectations.

Cápsula

Entrada típica

Saída típica

Status

video

Long video → dense LLM context.

[demo · 1h08 · 1080p]fonte

chapters + key quotes + on-screen actions · 2.4k tokenscompactado

89%multimodal

IML

image-llm

Screenshots/diagrams → in-context description (LLM add-on).

[dashboard.png · 1.4MB]fonte

KPIs read (OCR) + layout described · 380 tokenscompactado

94%multimodal

meeting

Audio → summary with decisions.

[client call · 47min]fonte

decisions, action items (owners+deadlines), next stepscompactado

ao vivomultimodal