parsec

78% fewer tokens and double your Claude Code limit

Saved by parsec users · as of 2026-09-09

$38,336

44B input tokens

Backed byANTLER

on SWE-bench Verified

Cheaper, faster, and better.

Cheaper

-39%

total cost · $89.65 vs $147.30

$1.45 per solved task — -44%, lowest of six arms.

Faster

10.8 h

vs 14.4 h — -25%

6.2 s mean latency, fastest arm. Woz was slowest: 17.8 h.

Better

62 / 100

solved vs 57 with no compression

Same model, same grader. Woz: 55.

Every Parsec number is graded on one fixed model and the official SWE-bench harness — the only arm that is both cheaper and faster than doing nothing. See the benchmark

How does parsec impact across the harness

Our curator model scores every chunk of tool output and thinking and drops information the agent never edits, cites, re-reads or reuses.

  • Search results45,804,275 tokens

    98.2%

  • Context & plans203,023,138 tokens

    84.9%

  • File reads336,629,352 tokens

    82.2%

  • Thinking71,210,142 tokens

    69.7%

  • Command output279,139,117 tokens

    68.1%

  • Web content2,805,491 tokens

    67.9%

  • Your prompts & the agent's answersnever touched

    0%

All trim-eligible: 78.4% never used

Measured on 28,487 real agent traces.

SWE-bench Verified · no compression → Parsec

The model doesn't need to think about what it doesn't see.

Parsec drops the tool output and thinking the agent never uses before the model sees the turn. Less in context means less to reason about and less written back out — and it solves more tasks, not fewer.

In · input tokens

312.2M
144.8M

-54%

Out · output tokens

3.00M
1.71M

-43%

Time · wall clock

14.4 h
10.8 h

-25%

Solves · /100

57
62

+5

One real task · pydata/xarray-4356 · same fix both ways

2.12M in · 15.9K out · 40 calls · $1.02189K in · 1.9K out · 11 calls · $0.22

Capability matrix · Parsec vs the field

Every other layer rewrites, summarizes, or deletes what your agent reads.

Capabilities and SWE-bench Verified results, Parsec vs Caveman, Headroom, Woz and RTK
 ParsecCavemanHeadroomWozRTK
Compresses tool outputyesyesyesyesshell only
Compresses thinkingyes no no no no
Trims the tool roster every turnyes no no no no
Never alters readable messagesyes no no no no
Kept content byte-exact (no rewording)yes no no no no
Learned from real agent tracesyes no no no no
Measured on the wire, not estimatedyesprovider-rep.estimatedestimatedisolated
Graded on task solves (SWE-bench, /100)6258585554
Cuts output tokens−43%−30% *−2%−1%+8%
Cuts input tokens (cache-aware)−54%−19%+6%−35%+16%
Cost per solved task$1.45$2.05$3.66$2.33$3.07
Mean added latency per call6.2 s6.9 s8.0 s14.3 s9.3 s
Runs localyesyesyes noyes

* Caveman cuts output by rewriting the reply into caveman-speak, not by reducing reasoning. Token, cost, time and solve columns are the six-arm SWE-bench Verified run (same model, official grader). Competitor claims as published on their sites and READMEs, 2026-09-09.

The full comparison — input types, per-task costs, and every workload →

testimonials

What engineers say

I love this. In the past week I saved $2.5K in api credits and 6B in tokens. I blasted my plan and never hit the limit.

Nicholas SwaminathanFounderDasein

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

Runs local · your credentials · measured per request

Pricing

Pro

Start here

Plugin + local proxy

$40/mo Free

  • 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

Deployed in your cloud

Talk to us

  • The whole stack — scoring API included — runs in your VPC
  • Licensed checkpoints, containerized
  • Fine-tuning on your traces, without them leaving your network

questions

FAQ

Which coding agents does it work with?

Claude Code, OpenAI Codex CLI, and opencode — one local proxy serves all three, and every ledger row records which tool it came from. Claude Code gets the full plugin surface: status line, hooks, and skills. Codex routes through parsec with your existing ChatGPT sign-in (or an API key), and opencode works with Anthropic API-key providers. One honesty caveat: OpenAI has no count_tokens endpoint, so Codex rows record actually-billed usage without the measured-savings counterfactual — we do not estimate what we cannot measure.

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, tokens are sent to us but never stored. The one exception is the hosted Team gateway, where you explicitly choose to run the proxy in our cloud with a BYOK API key. On Enterprise the direction reverses: the scoring API is deployed inside your own cloud, so no chunk text reaches us at all.

Can I use my Claude subscription (Max plan) or ChatGPT plan?

Yes. Traffic leaves your machine with your own OAuth token, same as stock Claude Code — the rewriting happens locally. Codex works the same way with your ChatGPT sign-in. 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.