Half your PDFs don't need OCR — pdf-inspector tells you which half
Most pipelines run every PDF through OCR, even though roughly 54% of them already contain real, extractable text. pdf-inspector is a Rust library that classifies a PDF as text-based, scanned, image-based, or mixed in about 10-50ms, so only pages that actually need OCR get routed there.
For text-based PDFs it extracts text with position awareness, preserves multi-column reading order, detects tables, and converts straight to clean Markdown — no OCR involved. On a 200-PDF benchmark it beat every other local parser on overall score, finishing the full run in under half a second.
It ships as a Rust crate and CLI, with native bindings for Python, Node.js, and browser WebAssembly. Install via
https://github.com/firecrawl/pdf-inspector
Most pipelines run every PDF through OCR, even though roughly 54% of them already contain real, extractable text. pdf-inspector is a Rust library that classifies a PDF as text-based, scanned, image-based, or mixed in about 10-50ms, so only pages that actually need OCR get routed there.
For text-based PDFs it extracts text with position awareness, preserves multi-column reading order, detects tables, and converts straight to clean Markdown — no OCR involved. On a 200-PDF benchmark it beat every other local parser on overall score, finishing the full run in under half a second.
It ships as a Rust crate and CLI, with native bindings for Python, Node.js, and browser WebAssembly. Install via
cargo add pdf-inspector, pip install pdf-inspector, or npm install @firecrawl/pdf-inspector, then call process_pdf().https://github.com/firecrawl/pdf-inspector
The Python engine that draws 3Blue1Brown's math animations
Manim turns code into precise, moving diagrams instead of static slides and formulas. You describe objects and transformations in Python, and it renders the animation frame by frame — built for the kind of math visuals that make a concept click instead of just sitting on a slide.
This is ManimGL, the original engine behind 3Blue1Brown's videos, distinct from the community fork aimed at broader stability. It runs on Python 3.10+, needs FFmpeg and OpenGL, and LaTeX if you want typeset equations in the scene.
Getting started is one command:
https://github.com/3b1b/manim
Manim turns code into precise, moving diagrams instead of static slides and formulas. You describe objects and transformations in Python, and it renders the animation frame by frame — built for the kind of math visuals that make a concept click instead of just sitting on a slide.
This is ManimGL, the original engine behind 3Blue1Brown's videos, distinct from the community fork aimed at broader stability. It runs on Python 3.10+, needs FFmpeg and OpenGL, and LaTeX if you want typeset equations in the scene.
Getting started is one command:
pip install manimgl. Clone the repo, run the included example scenes, and a window pops up playing your first rendered animation — from there it's your own Python classes driving the camera and the math.https://github.com/3b1b/manim
Your AI agent just met the laziest senior dev on the team
You ask for a date picker. Your agent installs a library, writes a wrapper component, and opens a debate about timezones. Ponytail stops that before it starts — the same instinct as a senior dev who looks at fifty new lines and quietly replaces them with one.
Before writing code it climbs a ladder: does this need to exist, is it already in the codebase, does the stdlib or platform already do it, can one line do it — only then it writes the minimum that works, after reading the surrounding code first. Validation, security, and accessibility are never cut.
On real Claude Code sessions editing a FastAPI + React repo: about 54% less code, 20% cheaper, 27% faster, same safety guardrails. Install as a plugin for Claude Code, Codex, or Copilot CLI.
…
You ask for a date picker. Your agent installs a library, writes a wrapper component, and opens a debate about timezones. Ponytail stops that before it starts — the same instinct as a senior dev who looks at fifty new lines and quietly replaces them with one.
Before writing code it climbs a ladder: does this need to exist, is it already in the codebase, does the stdlib or platform already do it, can one line do it — only then it writes the minimum that works, after reading the surrounding code first. Validation, security, and accessibility are never cut.
On real Claude Code sessions editing a FastAPI + React repo: about 54% less code, 20% cheaper, 27% faster, same safety guardrails. Install as a plugin for Claude Code, Codex, or Copilot CLI.
…
Stop hunting for which port your dev server picked this time
Every restart hands you a fresh port, a broken bookmark, a cookie that no longer matches. portless swaps that churn for one stable, named URL: run
It sets up HTTPS with HTTP/2 automatically, generating and trusting a local CA on first run, and it injects the right port flag even into frameworks like Vite or Astro that ignore the
It's a drop-in wrapper: swap
https://github.com/vercel-labs/portless
Every restart hands you a fresh port, a broken bookmark, a cookie that no longer matches. portless swaps that churn for one stable, named URL: run
portless myapp next dev and your app lives at https://myapp.localhost, every time, no matter what port the framework actually bound.It sets up HTTPS with HTTP/2 automatically, generating and trusting a local CA on first run, and it injects the right port flag even into frameworks like Vite or Astro that ignore the
PORT env var. Subdomains work out of the box too, so api.myapp and docs.myapp can run side by side, and git worktrees get their own branch-prefixed URL with zero config.It's a drop-in wrapper: swap
"next dev" for "portless run next dev" in package.json, or install it globally and just type portless. Works the same for a single app or a whole monorepo, and stays out of the way of anyone on the team who isn't using it.https://github.com/vercel-labs/portless
Your AI agent rambles. You pay per token for the rambling.
Caveman is a skill for Claude Code (and 30+ other agents) that tells your agent to talk like a caveman: short, blunt, no throat-clearing. The diagnosis survives, the fix survives, the code survives — only the filler goes. Across ten real coding prompts it cut output tokens by 65% on average, up to 87% on the best case.
It never touches code, commands, file paths, or error messages, only the prose wrapped around them. Written in Go, MIT licensed, and there's an optional proxy that shrinks what the agent reads too, for the sessions where input tokens are the real bill.
Install is one command:
Then type
…
Caveman is a skill for Claude Code (and 30+ other agents) that tells your agent to talk like a caveman: short, blunt, no throat-clearing. The diagnosis survives, the fix survives, the code survives — only the filler goes. Across ten real coding prompts it cut output tokens by 65% on average, up to 87% on the best case.
It never touches code, commands, file paths, or error messages, only the prose wrapped around them. Written in Go, MIT licensed, and there's an optional proxy that shrinks what the agent reads too, for the sessions where input tokens are the real bill.
Install is one command:
npx skills add JuliusBrussee/cavemanThen type
/caveman if your agent doesn't wake up on its own. Same brain, fewer words, smaller invoice.…
Your writing has AI tells. This tool strips them out.
Em dashes, hollow words like "testament," forced groups of three — readers spot AI-generated prose in seconds, and it undercuts writing that would otherwise land fine. Humanizer rewrites the text so it reads like a person wrote it, without changing what it says.
It works from 35 documented patterns in Wikipedia's "Signs of AI writing" project: inflated importance, sales language, fake-candid openings, curly quotes, and more. It drafts a first rewrite, checks it against the patterns and the original claims, then shows you both the draft and a short critique before the final version. Facts, code, data, and links are left alone. Give it a writing sample and it will match your voice instead of using its defaults.
Humanizer is a Markdown skill, so it runs in any agent that supports skills. Paste text and ask to humanize it, or point it at a file to rewrite the prose in place.
https://github.com/blader/humanizer
Em dashes, hollow words like "testament," forced groups of three — readers spot AI-generated prose in seconds, and it undercuts writing that would otherwise land fine. Humanizer rewrites the text so it reads like a person wrote it, without changing what it says.
It works from 35 documented patterns in Wikipedia's "Signs of AI writing" project: inflated importance, sales language, fake-candid openings, curly quotes, and more. It drafts a first rewrite, checks it against the patterns and the original claims, then shows you both the draft and a short critique before the final version. Facts, code, data, and links are left alone. Give it a writing sample and it will match your voice instead of using its defaults.
Humanizer is a Markdown skill, so it runs in any agent that supports skills. Paste text and ask to humanize it, or point it at a file to rewrite the prose in place.
https://github.com/blader/humanizer
An AI agent that remembers you, not just your last message
Most agents forget everything the moment you close the tab, so you re-explain your setup and preferences every single time. Hermes Agent keeps a closed learning loop instead: it builds skills from what it does, keeps refining them with use, and searches its own past sessions to pull back context on its own.
It isn't chained to a chat window either. Run it in a full terminal UI, or reach it from Telegram, Discord, Slack, WhatsApp, or Signal while it keeps working on a cloud VM, a $5 server, or serverless infrastructure that idles for pennies. Swap the underlying model anytime with a single command — Nous Portal, OpenRouter, OpenAI, or your own endpoint, no code changes.
It's written in Python and MIT-licensed, with one install script for Linux, macOS, WSL2, Termux, and native Windows.
…
Most agents forget everything the moment you close the tab, so you re-explain your setup and preferences every single time. Hermes Agent keeps a closed learning loop instead: it builds skills from what it does, keeps refining them with use, and searches its own past sessions to pull back context on its own.
It isn't chained to a chat window either. Run it in a full terminal UI, or reach it from Telegram, Discord, Slack, WhatsApp, or Signal while it keeps working on a cloud VM, a $5 server, or serverless infrastructure that idles for pennies. Swap the underlying model anytime with a single command — Nous Portal, OpenRouter, OpenAI, or your own endpoint, no code changes.
It's written in Python and MIT-licensed, with one install script for Linux, macOS, WSL2, Termux, and native Windows.
…
One Google model now forecasts almost any time series, zero-shot
TimesFM is a pretrained foundation model for time-series forecasting from Google Research. Point it at new data and it forecasts directly, no per-dataset training required.
Version 3.0 adds native multivariate forecasting: several related series, plus past and future covariates, produce one joint forecast. Output includes point predictions and nine quantiles for calibrated uncertainty. It ranks first on fev-bench, TIME, and GIFT-Eval, and already runs inside BigQuery ML and Google Sheets.
Install with
github.com/google-research/timesfm
TimesFM is a pretrained foundation model for time-series forecasting from Google Research. Point it at new data and it forecasts directly, no per-dataset training required.
Version 3.0 adds native multivariate forecasting: several related series, plus past and future covariates, produce one joint forecast. Output includes point predictions and nine quantiles for calibrated uncertainty. It ranks first on fev-bench, TIME, and GIFT-Eval, and already runs inside BigQuery ML and Google Sheets.
Install with
pip install timesfm[torch], load the 3.0 checkpoint from Hugging Face, and call predict on your own arrays. Code is Apache-2.0; the 3.0 pretrained weights are non-commercial only.github.com/google-research/timesfm
Run Claude Code, Codex, or Cline on local models — no keys, no bill
Cloud agents rack up token costs, API keys, and rate limits. Setting up local models yourself means guessing which quant fits your machine and hoping it's fast enough to use.
Magnitude is an open source inference server that ends the guessing. It profiles your hardware, recommends models that fit, then downloads and tunes them — decoding, concurrency, all set automatically. Models load on demand and unload when idle.
It plugs into the agent you already use: Pi, OpenCode, Hermes, OpenClaw, Codex, Claude Code, Oh My Pi, or Cline. One prompt walks your agent through setup, then everything runs fully offline — nothing leaves your machine.
Install with
https://github.com/magnitudedev/magnitude
Cloud agents rack up token costs, API keys, and rate limits. Setting up local models yourself means guessing which quant fits your machine and hoping it's fast enough to use.
Magnitude is an open source inference server that ends the guessing. It profiles your hardware, recommends models that fit, then downloads and tunes them — decoding, concurrency, all set automatically. Models load on demand and unload when idle.
It plugs into the agent you already use: Pi, OpenCode, Hermes, OpenClaw, Codex, Claude Code, Oh My Pi, or Cline. One prompt walks your agent through setup, then everything runs fully offline — nothing leaves your machine.
Install with
npm i -g @magnitudedev/cli, then run magnitude docs onboarding. Apache 2.0, written in TypeScript.https://github.com/magnitudedev/magnitude
❤1
One repo, dozens of unreported exploits, and the CVE credit is up for grabs
Public PoCs and vulnerability writeups usually rot across dozens of one-off repos that go dark the moment their author moves on. Exploitarium fixes that by folding every standalone PoC into one growing Python archive with a consistent structure per folder.
Scroll the table of contents and the target list reads like a greatest-hits round: Firefox, Ghidra, Docker, OpenSSH, curl, Redis, and dozens more. Some folders hold a single bug, others track whole families of variations on it.
Every proof of concept here is hand-typed by the author, not generated, and none of it had been publicly reported at the time of posting — find one first and the report, and any CVE that follows, is yours to claim. It's shared in good faith, to pull more people into vulnerability research rather than to arm anyone.
Clone it, pick a folder, read the writeup.
github.com/bikini/exploitarium
Public PoCs and vulnerability writeups usually rot across dozens of one-off repos that go dark the moment their author moves on. Exploitarium fixes that by folding every standalone PoC into one growing Python archive with a consistent structure per folder.
Scroll the table of contents and the target list reads like a greatest-hits round: Firefox, Ghidra, Docker, OpenSSH, curl, Redis, and dozens more. Some folders hold a single bug, others track whole families of variations on it.
Every proof of concept here is hand-typed by the author, not generated, and none of it had been publicly reported at the time of posting — find one first and the report, and any CVE that follows, is yours to claim. It's shared in good faith, to pull more people into vulnerability research rather than to arm anyone.
Clone it, pick a folder, read the writeup.
github.com/bikini/exploitarium
38 diagram types that don't look like Mermaid vomited on your README
Ask an AI coding assistant for a diagram and you get the same thing every time: generic rounded boxes, a default color scheme, zero relation to your actual brand. Fixing it by hand means opening Figma and losing thirty minutes to a color picker.
This is a skill for Claude Code, Codex, and Pi that replaces the whole habit. It ships 38 editorial diagram types — architecture, flowcharts, state machines, pyramids, and more — as self-contained HTML and SVG. No build step, no JavaScript, no external image dependencies, and no shadows or Mermaid-style clutter.
Each diagram comes in three static variants — minimal light, minimal dark, full editorial — and the skill can read your website to match colors and style automatically. Install it, ask for a diagram, open the HTML file in a browser.
github.com/cathrynlavery/diagram-design
Ask an AI coding assistant for a diagram and you get the same thing every time: generic rounded boxes, a default color scheme, zero relation to your actual brand. Fixing it by hand means opening Figma and losing thirty minutes to a color picker.
This is a skill for Claude Code, Codex, and Pi that replaces the whole habit. It ships 38 editorial diagram types — architecture, flowcharts, state machines, pyramids, and more — as self-contained HTML and SVG. No build step, no JavaScript, no external image dependencies, and no shadows or Mermaid-style clutter.
Each diagram comes in three static variants — minimal light, minimal dark, full editorial — and the skill can read your website to match colors and style automatically. Install it, ask for a diagram, open the HTML file in a browser.
github.com/cathrynlavery/diagram-design
One coding agent, any model, zero lock-in
Most coding assistants chain you to a single model vendor. OpenCode is a terminal-native coding agent written in TypeScript that lets you plug in whatever model you want and keeps it that way — free and open source, no rented intelligence.
It ships two built-in agents you flip between with a single Tab press: build mode edits files and runs commands freely, plan mode stays read-only and asks permission before touching your shell, which makes it safe for poking around a codebase you don't fully trust yet. A general-purpose subagent handles multistep searches in the background.
Install it with one command, point it at any repo, and start handing it real tasks — it reads, edits, and runs things right in your terminal, no IDE plugin required.
https://github.com/anomalyco/opencode
Most coding assistants chain you to a single model vendor. OpenCode is a terminal-native coding agent written in TypeScript that lets you plug in whatever model you want and keeps it that way — free and open source, no rented intelligence.
It ships two built-in agents you flip between with a single Tab press: build mode edits files and runs commands freely, plan mode stays read-only and asks permission before touching your shell, which makes it safe for poking around a codebase you don't fully trust yet. A general-purpose subagent handles multistep searches in the background.
Install it with one command, point it at any repo, and start handing it real tasks — it reads, edits, and runs things right in your terminal, no IDE plugin required.
https://github.com/anomalyco/opencode
Stop letting a bad format string crash your C++ program
{fmt} replaces printf and iostreams in C++ with something faster and harder to misuse. printf can crash on a mismatched format string, and iostreams force chains of << just to print a few values. {fmt} uses Python-style format strings checked at compile time, so an invalid specifier is caught before the program ever runs.
It's built for speed: a Dragonbox-based float formatter with correct rounding, minimal dynamic allocation, and an optional compile-time format path make it tens of percent to tens of times faster than sprintf and iostreams. It also formats containers, dates, times, and colored terminal output through the same simple API.
The core is a few headers with no external dependencies, MIT-licensed, and it's the basis for C++20's std::format and C++23's std::print. Drop it into an existing project and start formatting.
github.com/fmtlib/fmt
{fmt} replaces printf and iostreams in C++ with something faster and harder to misuse. printf can crash on a mismatched format string, and iostreams force chains of << just to print a few values. {fmt} uses Python-style format strings checked at compile time, so an invalid specifier is caught before the program ever runs.
It's built for speed: a Dragonbox-based float formatter with correct rounding, minimal dynamic allocation, and an optional compile-time format path make it tens of percent to tens of times faster than sprintf and iostreams. It also formats containers, dates, times, and colored terminal output through the same simple API.
The core is a few headers with no external dependencies, MIT-licensed, and it's the basis for C++20's std::format and C++23's std::print. Drop it into an existing project and start formatting.
github.com/fmtlib/fmt
Dictation that never phones home your voice
OpenWhispr turns speech into text at your cursor, in any app, with a single hotkey press. Run it fully offline using Whisper or NVIDIA Parakeet — your audio never leaves the device — or switch to cloud models with your own API key when you want more speed. No telemetry, no data collection, ever.
Past plain dictation it becomes a voice-driven assistant: talk to GPT-5, Claude, Gemini, Groq, or a local model, dictate in one language and get the text pasted in another, or let it transcribe meetings live with on-device speaker diarization and voice fingerprinting, no cloud required.
It's a JavaScript/Electron app for macOS, Windows, and Linux, fully open source. Grab a build from the releases page, or clone the repo and run
https://github.com/OpenWhispr/openwhispr
OpenWhispr turns speech into text at your cursor, in any app, with a single hotkey press. Run it fully offline using Whisper or NVIDIA Parakeet — your audio never leaves the device — or switch to cloud models with your own API key when you want more speed. No telemetry, no data collection, ever.
Past plain dictation it becomes a voice-driven assistant: talk to GPT-5, Claude, Gemini, Groq, or a local model, dictate in one language and get the text pasted in another, or let it transcribe meetings live with on-device speaker diarization and voice fingerprinting, no cloud required.
It's a JavaScript/Electron app for macOS, Windows, and Linux, fully open source. Grab a build from the releases page, or clone the repo and run
npm install && npm run dev to build it yourself.https://github.com/OpenWhispr/openwhispr
Your own hedge fund, run entirely by AI agents
Building a real trading desk normally takes analysts, risk managers, and traders working around the clock. AutoHedge replaces that team with a swarm of specialized agents: a Director for strategy, a Quant for analysis, a Risk Manager for position sizing, and an Execution Agent to place the trades.
Each agent hands its output to the next in a structured pipeline, so a thesis gets generated, checked, sized, and executed without a human in the loop. Everything comes back as JSON with full logging, so you can audit exactly why a trade happened.
It's fully open source and written in Python. Install with
https://github.com/The-Swarm-Corporation/AutoHedge
Building a real trading desk normally takes analysts, risk managers, and traders working around the clock. AutoHedge replaces that team with a swarm of specialized agents: a Director for strategy, a Quant for analysis, a Risk Manager for position sizing, and an Execution Agent to place the trades.
Each agent hands its output to the next in a structured pipeline, so a thesis gets generated, checked, sized, and executed without a human in the loop. Everything comes back as JSON with full logging, so you can audit exactly why a trade happened.
It's fully open source and written in Python. Install with
pip install -U autohedge, drop in your API keys and wallet, and it's already trading autonomously on Solana, with Coinbase support on the way.https://github.com/The-Swarm-Corporation/AutoHedge
A directory of tools that never ask for your email
FckSignups is a curated, open-source list of browser tools you can use the instant you click through — no account, no verification email, no credit card just to resize an image. Every entry is checked to work with zero signup.
The catch that isn't: over 200 tools, sorted into categories like design, development, privacy, and writing, each open source in its own right. Entries that stand out from the crowd get flagged as featured, so the list stays browsable instead of turning into a wall of duplicates.
It's a React + TypeScript app you can run yourself in four commands, or just browse it live. Missing your favorite no-signup tool? Submit it through the site or open an issue — the schema is simple: name, link, category, done.
…
FckSignups is a curated, open-source list of browser tools you can use the instant you click through — no account, no verification email, no credit card just to resize an image. Every entry is checked to work with zero signup.
The catch that isn't: over 200 tools, sorted into categories like design, development, privacy, and writing, each open source in its own right. Entries that stand out from the crowd get flagged as featured, so the list stays browsable instead of turning into a wall of duplicates.
It's a React + TypeScript app you can run yourself in four commands, or just browse it live. Missing your favorite no-signup tool? Submit it through the site or open an issue — the schema is simple: name, link, category, done.
…
A headless browser that isn't secretly Chrome in a trenchcoat
Lightpanda is built from scratch for AI agents and automation, no Chromium fork, no WebKit patch underneath. Written in Zig, it skips the weight that headless Chrome drags along: on a 100-page crawl it used 123MB of memory against Chrome's 2GB, and finished in 5 seconds instead of 46.
It plugs into what you already use: point Puppeteer or Playwright at its CDP server and nothing else in your script changes. There's also a WebDriver Bidi mode, and an `lightpanda agent` command that drives the browser from plain-English instructions in your terminal, then exports the session as a plain JavaScript script you can replay without an LLM.
Grab a nightly build via Homebrew, the AUR, Docker, or a direct binary for Linux and macOS, or build it from source.
…
Lightpanda is built from scratch for AI agents and automation, no Chromium fork, no WebKit patch underneath. Written in Zig, it skips the weight that headless Chrome drags along: on a 100-page crawl it used 123MB of memory against Chrome's 2GB, and finished in 5 seconds instead of 46.
It plugs into what you already use: point Puppeteer or Playwright at its CDP server and nothing else in your script changes. There's also a WebDriver Bidi mode, and an `lightpanda agent` command that drives the browser from plain-English instructions in your terminal, then exports the session as a plain JavaScript script you can replay without an LLM.
Grab a nightly build via Homebrew, the AUR, Docker, or a direct binary for Linux and macOS, or build it from source.
…
Your scraper just got a Cloudflare-proof disguise
Playwright gets blocked. Headless Chrome gets fingerprinted on the first request. Even stealth plugins end up being the tell that gives you away. camofox-browser sidesteps all of it by wrapping Camoufox, a Firefox fork that spoofs navigator properties, WebGL, AudioContext, and screen geometry at the C++ level — before any JavaScript ever runs to detect it.
It ships as a REST API built for agents rather than humans: accessibility snapshots instead of bloated HTML, stable element refs like
Try it with
…
Playwright gets blocked. Headless Chrome gets fingerprinted on the first request. Even stealth plugins end up being the tell that gives you away. camofox-browser sidesteps all of it by wrapping Camoufox, a Firefox fork that spoofs navigator properties, WebGL, AudioContext, and screen geometry at the C++ level — before any JavaScript ever runs to detect it.
It ships as a REST API built for agents rather than humans: accessibility snapshots instead of bloated HTML, stable element refs like
e1, e2 for reliable clicking, and search macros for sites like Google and Reddit. It's a drop-in replacement for Puppeteer or Playwright, idles at ~40MB of memory, and runs fine on a $5 VPS or a Raspberry Pi.Try it with
npx @askjo/camofox-browser, or clone the repo, run npm install && npm start, and hit http://localhost:9377. Docker and an OpenClaw plugin are both supported out of the box.…
Your AI agent can write HTML all day. Now it can ship video too.
HyperFrames is an open-source framework that renders HTML, CSS, media, and seekable animations straight into deterministic MP4 files. No editor, no timeline — just markup in, video out, the same result every time.
It's built for agents, not humans clicking timelines. Skills teach Claude Code, Cursor, Codex, Gemini CLI, and other agents the full loop: plan the video, write valid HTML, wire up animations, add media, lint, preview, render. One prompt describing a video is enough to trigger it.
Try it with
…
HyperFrames is an open-source framework that renders HTML, CSS, media, and seekable animations straight into deterministic MP4 files. No editor, no timeline — just markup in, video out, the same result every time.
It's built for agents, not humans clicking timelines. Skills teach Claude Code, Cursor, Codex, Gemini CLI, and other agents the full loop: plan the video, write valid HTML, wire up animations, add media, lint, preview, render. One prompt describing a video is enough to trigger it.
Try it with
npx skills add heygen-com/hyperframes, or for agent and non-interactive runs npx hyperframes skills update. Written in TypeScript, Apache-2.0 licensed.…
Your AI agent is burning its own context — this MCP server stops it
Every tool call dumps raw data into your agent's context window — a single Playwright snapshot alone costs 56 KB. After half an hour, 40% of that context can be gone, and when the conversation compacts, the agent forgets what it was doing.
Context Mode is an MCP server that sandboxes tool output before it reaches the model — 315 KB drops to 5.4 KB, a 98% cut — and persists session memory in SQLite with FTS5 search, so a resumed session picks up exactly where it left off.
It also changes the habit: instead of reading dozens of files into context, the agent writes a small script that does the work and logs only the result. Routing is enforced via hooks across 17 platforms, not left as something to remember.
Install through Claude Code's plugin marketplace or npm, then run the built-in doctor command to confirm everything is wired up.
mksglu/context-mode
Every tool call dumps raw data into your agent's context window — a single Playwright snapshot alone costs 56 KB. After half an hour, 40% of that context can be gone, and when the conversation compacts, the agent forgets what it was doing.
Context Mode is an MCP server that sandboxes tool output before it reaches the model — 315 KB drops to 5.4 KB, a 98% cut — and persists session memory in SQLite with FTS5 search, so a resumed session picks up exactly where it left off.
It also changes the habit: instead of reading dozens of files into context, the agent writes a small script that does the work and logs only the result. Routing is enforced via hooks across 17 platforms, not left as something to remember.
Install through Claude Code's plugin marketplace or npm, then run the built-in doctor command to confirm everything is wired up.
mksglu/context-mode