parsec

the context. ½ the cost.

a Claude Code plugin that compresses agent context per turn with a learned curator — model traffic goes straight from your machine to Anthropic, never through our cloud, and savings are measured per-request, never estimated.

claude plugin marketplace add https://github.com/daseinlabs/claude-plugins
read how it works

demo

install, run, ask for the number

two commands to install. one skill to see what you saved.

~ — claude code

claude plugin marketplace add https://github.com/daseinlabs/claude-plugins
marketplace added: parsec-marketplace
claude plugin install parsec@parsec-marketplace
parsec installed — prebuilt binary, no postinstall
claude
parsec active — local proxy running, requests go straight to Anthropic
/parsec:savings
12.4k input tokens not sent this session
across 31 requests — each measured via count_tokens

illustrative session — your numbers come from per-request count_tokens measurement, never an estimate.

how it works

local proxy, hosted brain

  1. install the plugin

    one command in Claude Code — no base-URL surgery, no wire config. agents, hooks, and the savings status line arrive as plugin surfaces.

  2. a local proxy curates each turn

    chunking, freezing, and splicing happen on your machine. the hosted brain returns keep/cut scores; the proxy applies the cut, and your request rides your own credentials to Anthropic.

  3. savings are measured, not modeled

    every request runs a free count_tokens probe of the original body against what was actually billed. no modeled baseline, no extrapolation — if the ledger is empty, the report says so.

what you get

a learned curator, with receipts

  • learned curator

    a GNN scores every chunk of context, every turn. the proxy applies keep/cut locally. the model sees what matters.

  • exploration mapsfree tier

    a parsec:explore agent maps the repo up front. the agent stops re-discovering the same files every session.

  • no-reread hookfree tier

    re-reading a file already in context gets denied with a pointer to where it lives. a loop-breaker catches the same command run on repeat.

  • honest measurement

    every request runs a count_tokens probe of the original body. savings are measured against what you were actually billed — never modeled.

  • data plane local

    model traffic leaves your machine, with your credentials. subscription tokens never touch our cloud.

  • fail open

    every layer degrades to passthrough on error. fail-opens are counted and visible, not hidden.

the product

what it looks like

a ledger, not a chart deck: dollars saved per request on the dashboard, and the same accounting surfaced in your session.

app.getparsec.ai

usage & savings

Total cost saved
$7.41
1.4M input tokens not sent · measured on 391 of 402 requests
Spend
$18.63
billed at list price
Fail-open
0
served passthrough

By model

ModelRequestsTokens savedCost saved
claude-opus-5118486k$5.12
claude-sonnet-5284913k$2.29
the savings ledger — measured per request
claude

Read src/lib/platform.ts

re-read denied: file already in context (insist once to override)

parsec 12.4k saved

the plugin surface — guardrails and savings in-session

illustrative — rendered UI, not screenshots; numbers shown are examples.

measurement honesty

measured, never modeled

before curation touches anything, the proxy calls Anthropic's free count_tokens on the original request body — the exact bytes you would have sent.

after the response comes back, it records what you were actually billed. the savings number is the difference between the two, computed per request.

every row lands in a local ledger — ~/.parsec/ledger.jsonl — that you can open and read yourself.

if the probe fails, the row records a hole — never an invented number.

{
  "counterfactual_input_tokens": 21968,
  "billed_input_tokens": 1204,
  "cache_read_tokens": 9861,
  "model": "claude-sonnet-5",
  "request_id": "req_011CQx…",
  "fail_open": false
}
one row of the local savings ledger — illustrative values.

testimonials

what engineers say

By turn thirty our agents were dragging the entire transcript into every request. parsec was the first tool that trimmed it without asking me to trust a summary I couldn't inspect.

Mara Okonkwostaff engineerDraywick Labs

The ledger is what convinced me. Every request shows the counterfactual next to what was actually sent — I didn't have to believe a dashboard, I could read the receipts.

Anders Viethinfrastructure leadLarkspur Systems

My Max plan used to hit the weekly wall by Thursday. Long agent sessions just keep going now — same model, same credentials, nothing routed through anyone else's cloud.

Priya Raghunathagent platform engineerNinebark Automation

tiers

start free. the learned curator is the upgrade.

the free plugin is a real product, not a demo. paid tiers add the brain — the part that stays server-side.

free

start here

the plugin

free

  • exploration maps
  • no-reread hook
  • savings status line
  • runs on your plan or key — no account needed

pro

plugin + local proxy

coming soon

  • the full learned curator via the hosted scoring API
  • tool-schema prune
  • governor

team

hosted BYOK gateway in our cloud

coming soon

  • API keys only — never subscription tokens
  • zero local install
  • centralized reporting

enterprise

self-host

talk to us

  • licensed checkpoints
  • containerized stack
  • on-prem fine-tuning on your traces

on subscription plans there is no per-token bill — the win is more Claude Code inside the same rate limits.

faq

straight answers

does my code go to your servers?

on the pro tier, the local proxy sends chunk text and structural features to our scoring API over TLS and gets keep/cut scores back — your model requests never route through us; they leave your machine with your own credentials. on the free tier, nothing is sent to us at all. the one exception is the hosted team gateway, where you explicitly choose to run the proxy in our cloud with a BYOK API key.

can I use my Claude subscription (Max plan)?

yes. traffic leaves your machine with your own OAuth token, same as stock Claude Code — the rewriting happens locally. we never route subscription tokens through our cloud; the hosted gateway is BYOK API keys only.

what happens when your API is down?

the proxy fails open. your request passes through uncompressed, so a scoring outage never blocks a session. fail-open events are counted and visible, not silently swallowed.

how are savings calculated?

per request, the proxy runs a count_tokens probe on the original body and compares it against what was actually billed. that measured counterfactual is the only source of savings numbers — never a modeled baseline. the ledger is kept locally, alongside your traffic.

does it train on my data?

no by default. telemetry is opt-in and tiered, and the default is off. you can preview exactly what would upload before opting in, and purge requests are honored.

is it open source?

the client stack — plugin, proxy, engine, mapgen, contracts — is MIT. the trained model and the training stack are not. we open-source the deterministic shell and sell the brain.

get started

two commands from here

claude plugin marketplace add https://github.com/daseinlabs/claude-plugins

claude plugin install parsec@parsec-marketplace

the free tier needs no account — exploration maps, the no-reread hook, and the savings ledger run from the plugin alone.