furl-ctx 1.1.0

The context compression layer for AI agents - typically 0-54% fewer tokens on real high-entropy content, up to 95% on repetitive logs/fixtures (best-case ceiling), reversible via CCR

  ███████╗ ██╗   ██╗ ██████╗  ██╗
  ██╔════╝ ██║   ██║ ██╔══██╗ ██║
  █████╗   ██║   ██║ ██████╔╝ ██║
  ██╔══╝   ██║   ██║ ██╔══██╗ ██║
   ██║      ╚██████╔╝ ██║  ██║ ███████╗
   ╚═╝       ╚═════╝  ╚═╝  ╚═╝ ╚══════╝
 The context compression layer for AI agents
  

Typically 0–54% fewer tokens on real high-entropy content · up to 95% on repetitive logs/fixtures (honest read) · a Claude Code plugin · local-first · reversible

Install · Proof · Library · Community

---

Furl compresses everything your Claude Code agent reads — tool outputs, logs, web fetches, RAG chunks — before it reaches the model. Same answers, a fraction of the tokens. Every dropped byte stays retrievable on demand.

The name is nautical: to furl is to roll up a sail — Furl rolls long context up out of the model's way and keeps it on a line, ready to unfurl (retrieve) the instant you need it.

Furl was extracted from the author's Headroom context-engineering experimentation project — the early commit history carries that lineage.

Install

Two commands inside Claude Code:

/plugin marketplace add omar-y-abdi/furl
/plugin install furl@furl

That's it — this installs the compression hook, the MCP tools, and the skill. No pip install, no setup: Furl fetches itself on first use. Requires uv on your PATH (same as the official serena plugin).

What you get

• Auto-compression hook — shrinks large Bash / WebFetch / WebSearch / Task (sub-agent) outputs before they enter context. Fail-open: never breaks a tool call. It does not touch your Read / Grep / Glob file reads — by design, so a later Edit still sees exact file bytes; those reads (often a coding agent's largest context cost) pass through uncompressed (why). One honest limit: when an output is so large that Claude Code itself persists it to a file and hands the model only a file reference, there is no inline output for the hook to compress.
• Signal-aware offload + sliceable retrieval — a payload too big to compress inline (e.g. a 33 MB trace) comes back as a structured summary (schema, per-field value histograms, example rows) instead of a truncated head/tail, and the agent pulls a narrow slice on demand — retrieve(hash, select_field="name", select_equals="DroppedFrame") or a numeric range — without materializing the whole thing.
• MCP tools — furl_compress, furl_retrieve, furl_stats, furl_purge (erase stored originals), furl_search (find by content substring), furl_list (list stored entries). A seventh tool, furl_read, exists but is off by default — enable with FURL_MCP_READ=1 (see LIBRARY.md).
• Skill — explains the <<ccr:HASH>> retrieval flow and how to tune or disable it.

Tuning, disabling (FURL_HOOK_ENABLED=0), and the full reference: plugins/furl/README.md. Retrieval TTL differs by surface: the library defaults to 30 minutes; this Claude Code plugin ships a 24 h window (FURL_CCR_TTL_SECONDS=86400) governing both the hook's offloads and the MCP tools' stores; the furl CLI defaults to the same 24 h. (A bare MCP server without a valid FURL_CCR_TTL_SECONDS keeps a 1 h session TTL for its tool-stored entries, while dropped-row originals embedded in compressed output follow the library's 30-minute default — the full 24 h window needs the env set, as the plugin ships it.)

A note on version numbers: the Claude Code plugin versions independently from the furl-ctx engine it pins — a plugin release doesn't always mean an engine release, and vice versa. /plugin shows the plugin version; GitHub Releases and CHANGELOG.md track the engine version; the SessionStart banner shows both together (furl <plugin> · engine furl-ctx <engine>), which is the quickest way to see both numbers at once.

Proof

Token reduction on real captured data — a dated snapshot (inputs committed under benchmarks/data/ for auditability; a re-run measures the current engine, so absolute counts can drift from this table — the honest-read band below is the authoritative check). Every number uses the engine's own tokenizer; needle recall is 100% (a known unique row is always recoverable, in the output or via CCR). Read every figure below as a best-case ceiling, not a typical — the honest read follows.

Best-case ceilings — low-entropy dev fixtures (the compressor's happy path):

Dataset	Items	Before	After	Reduction	Info retention
code	7	41,025	471	99%	100%
multiturn	135	14,866	2,073	86%	100%
logs	90	8,595	619	93%	100%
search	90	4,102	318	92%	100%
repeated logs	90	3,621	120	97%	100%
disk	9	694	279	60%	100%

Across the corpus: 95% fewer tokens (72,903 → 3,880) at 100% information retention. Full methodology and the 6-seed adversarial sweep: BENCHMARKS.md.

Honest read: the numbers above are best-case, low-entropy ceilings measured on the dev fixtures — two independent, out-of-sample audits show they degrade by 6–43pp on fresh high-entropy / near-unique / realistic data (exactly where real logs and listings live). On genuinely high-entropy content, honest lossless savings sit in the 0–54% band, not 60–95% (code 0%, search 40%, repeated_logs 54%); read every figure here as a ceiling, not a typical, and see the tier-aware breakdown in BENCHMARKS.md.

The code row's 99% is CCR-offload of a large non-file-read tool output (e.g. Bash dumping source text); an agent's own Read/Grep/Glob file access bypasses the compression hook by design and passes through unchanged, at 0%.

Also a Python library

The same engine drops into any Python app or MCP host: from furl_ctx import compress. Install, usage, pipeline internals, prompt-caching contract, and the full FURL_* config reference live in LIBRARY.md.

Community

Questions or bug reports → open a GitHub issue (the surest way to reach the maintainer). For chat, there's a Discord.

Maintainer note: Furl is solo-maintained today — one person handles issues, PRs, and security reports, so response times vary with availability. CONTRIBUTING.md covers how PRs get reviewed and SECURITY.md covers the vulnerability-disclosure process; both hold regardless of team size.

License

Apache 2.0 — see LICENSE.