git:20260807.6a54034 to git:20260914.82b4c8f

25 added, 21 removed. Audit A to A.

- # furl

- furl is context compression, prompt compression, and token optimization for AI agents. It folds repeated tool output into a small hash-addressed marker, keeps the lines that matter so an agent spends fewer tokens, and returns any original the moment the agent asks, byte-exact for raw text and a semantically-complete re-serialization for structured JSON arrays. furl ships as plugins for Claude Code and Codex plus a standalone MCP server, and nothing is ever thrown away.

- ## Six real captures

- These numbers are real furl 1.2.0 output, generated by site/data/generate.py.

- - Application logs: 6,213 tokens to 263 tokens, 95.8 percent fewer, router:search

- - Crash report: 2,445 tokens to 302 tokens, 87.6 percent fewer, router:ccr_offload

- - JSON API response: 3,791 tokens to 218 tokens, 94.2 percent fewer, router:smart_crusher

- - Test output: 2,745 tokens to 283 tokens, 89.7 percent fewer, router:ccr_offload

- - CI build log: 2,400 tokens to 307 tokens, 87.2 percent fewer, router:ccr_offload

- - Query result: 5,036 tokens to 549 tokens, 89.1 percent fewer, router:search

+ # Furl

- Across all six, 22,630 tokens in and 1,922 out, which is 91.5 percent fewer.

+ > Context compression for AI agents, with stored details retrievable through CCR.

- ## Honest bounds

+ The current package is 1.4.0. This website replays six historical synthetic examples processed

+ with Furl 1.2.0 on 2026-07-13; it does not run live compression or accept uploads.

+ Recorded aggregate: 22,630 tokens before, 1,922 after (91.5% saved) on repetitive

+ generated machine output. These are best-case examples, not a guarantee for new inputs.

+ See BENCHMARKS.md in the repository for methods and the broader 0–54% range.

+ Raw stored text can be retrieved byte-exact. Structured JSON may be re-serialized.

+ Retention, eviction, configured redaction and storage availability apply. Never

+ promise indefinite retrieval, guaranteed savings or automatic ChatGPT hooks.

- furl folds only what repeats. On repetitive machine output the best case reaches up to about 95 percent fewer tokens. On high-entropy prose the honest range is lower, roughly 0 to 54 percent. Hands-off automatic compression works on Claude Code 2.1.163 and newer; the PostToolUse hook mirrors each replacement to the tool's output shape, so the harness honors it. It was built for anthropics/claude-code#68951. The manual API and the MCP tools work on every version and lose nothing. The 95 percent ceiling, the 91.5 percent six-capture figure above, and this 0 to 54 percent band are reconciled in one place: BENCHMARKS.md, the section The three public headline numbers.

+ ## Website

+ - [Demo](https://furl-ctx.vercel.app/)

+ - [Connect](https://furl-ctx.vercel.app/connect)

+ - [Support](https://furl-ctx.vercel.app/support)

+ - [Privacy](https://furl-ctx.vercel.app/privacy)

+ - [Cookies](https://furl-ctx.vercel.app/cookies)

+ - [Terms](https://furl-ctx.vercel.app/terms)

- ## Links

+ ## Source of truth

+ - [Repository](https://github.com/omar-y-abdi/furl-ctx)

+ - [Benchmarks](https://github.com/omar-y-abdi/furl-ctx/blob/main/BENCHMARKS.md)

+ - [Retention](https://github.com/omar-y-abdi/furl-ctx/blob/main/CCR-RETENTION.md)

+ - [PyPI](https://pypi.org/project/furl-ctx/)

- - Site: https://furl-ctx.vercel.app/

- - GitHub: https://github.com/omar-y-abdi/furl-ctx

- - PyPI: https://pypi.org/project/furl-ctx/

+ The local stdio server can connect through a private Secure MCP Tunnel. The

+ static website is not a public MCP endpoint. Public HTTP deployments must be

+ separately configured with OAuth and durable isolated per-user storage.