Skip to content
trimwire

What is trimwire?

trimwire is a tiny local gateway that prunes Claude Code’s API context on every request. Claude Code re-sends the whole conversation on every turn, so long sessions fill up with stale tool output, superseded file reads, and old logs. trimwire sits between Claude Code and Anthropic and removes that dead weight before it’s sent — keeping the current task from being buried in history.

Claude Code  ──►  trimwire (prune on every request)  ──►  Anthropic API
                  • deterministic, cache-safe
                  • optional summarizer
                  • leaves re-read cues for what it elided

It’s a transparent proxy: same auth, same sampling, and — on the default path — the same system prompt (the opt-in system_shape_normalize strategy is the only exception); only the conversation context changes (that’s the point). It works with API key, Pro, and Max (API key is clearly fine; Pro/Max OAuth has a grey-area ToS note — see FAQ.md).

Why it matters

Section titled “Why it matters”

When a session outgrows the model’s window, some context has to be dropped — and any lossy step (a model summarizer, including trimwire’s own, or a plain window cutoff) can discard older detail. trimwire’s mechanical difference: it prunes only what is structurally redundant or older than a configured window, and leaves a small [trimwire: …] marker where it elided content — a retrieval cue the agent can act on by re-reading the source. You get a cleaner, more focused request, and the markers can help the agent re-read an elided source when that source is still available as a file or re-runnable tool.

Section titled “Recommended defaults”
  • Pruning profile: default (ON). Eight cache-safe strategies, aggressive knobs — the shipped default. Recommended for almost everyone.
  • gentle is a lighter-touch, lower-savings profile — it prunes less; use it only if you specifically want minimal pruning. It is not a “recall mode” — see the recall note below.
  • Summarizer: off by default (engine = "model-free"). Pruning is purely deterministic out of the box. The optional summarizer compresses old content with a model; when enabled, the recommended backend is local qwen3.5:4b (via ollama — no API key, nothing leaves your machine). It is never load-bearing: any failure falls back to deterministic pruning. See SUMMARIZER.md.
  • Need an old detail back? Keep default on and ask the agent to re-read the file or re-run the tool — see the recall question in FAQ.md.

What the numbers look like

Section titled “What the numbers look like”

Request-size reduction depends entirely on session shape and on the mode; trimwire does nothing when there’s no redundancy. The figures below are model-free pruning, default profile — offline replay through the real strategy code (reproducible via cargo run --release --example bench) — spanning the full range:

  • ~0% on plain chat with no tools (no-op).
  • ~60–95% on tool-, read-, and browser-heavy sessions (the common agentic case).
  • Savings grow with session length — the longer and more repetitive the session, the more there is to prune.

The gentle profile prunes much less (0% on most of those corpora; it only fires on dedup/duplication-heavy shapes), and the optional summarizer is a separate lever (off by default). Per-mode numbers are in RESULTS.md.

Cost is a side effect and non-monotonic (prompt-cache hits bill at ~0.1×, so pruning old content can bust the cache): short sessions are a wash-to-slight-loss, long ones win (about −55% cache-weighted cost at 256 turns in our benchmark cost model, model-free default). The optional summarizer helps most on long sessions (up to roughly −65% cache-weighted cost observed on one long 981-turn session — a summarizer-mode best case, offline, not a guarantee). Overhead is roughly sub-2 ms per request (host-dependent), off the network path. Full tables, profiles, and the cost model live in benchmark/; for figures from your traffic, run trimwire stats.

The figures above are offline benchmark/cost-model replay. RESULTS.md keeps the live claude -p measurements, the reproducible benchmark, and the offline cost-model numbers in three clearly-separated buckets — including the honest typical-session live figure (~17% on a real session) versus the high percentages from adversarial read-heavy fixtures.

Limits & honest caveats

Section titled “Limits & honest caveats”
  • Cost can go up on short sessions (cache busting). trimwire reports headroom, not a guaranteed dollar figure.
  • The “focus ratio” is a byte-share proxy. That a cleaner context improves the model’s output is plausible but not proven; trimwire makes no quality-lift claim.
  • Very large / multi-hundred-million-token figures are projections only — no model accepts that much context in one request, so they can’t be measured directly.
  • Model-compatibility ceilings above ~128 KB are directional (small-N) — verify with a probe before relying on them. See MODEL-COMPATIBILITY.md.

Where to go next

Section titled “Where to go next”