What it does

All text below is quoted directly from the project README.

>One MCP for reverse engineering across binaries, applications, and runtime behavior.

See a feature you like. Understand how it works, down to the binary level.

See a feature in an app that you want in your own product? Ask your agent to investigate it with REA. It can inspect the app without its source code, explain how the feature works, show the evidence, and build a version for your project.

REA connects your agent to tools for inspecting native binaries, JavaScript and Electron apps, .NET assemblies, and websites. You can also use the same tools from your terminal. Analysis runs locally, and results include the evidence and limitations behind each conclusion.

What you can analyze:

• Native binaries: Pseudocode, assembly, strings, symbols, calls and references
• Offline ELF layout: Sections, segments, original symbols/relocations and static mitigation candidates
• JavaScript / Electron: Modules, imports, source maps, routes, IPC and native add-on relationships
• Websites: Page structure, scripts, network observations and requested screenshots
• Saved network captures: Requests, responses, exposed payloads and source locations
• .NET assemblies: Metadata, CIL instructions, declared native dependencies and build comparisons
• Android APKs: Manifest declarations, classes, decompiled methods and references
• Firmware: Regions, extraction results and native-analysis handoffs
• Packages and resources: File inventories, digests, plists, Apple bundle anatomy and extracted resources
• Process behavior: Terminal output, interactions, exit and filesystem observations, and run comparisons

>With Node.js and npm installed, run: npx rea-agents setup

Repository

https://github.com/morluto/rea

Repository stats

Metric	Value
Stars	17,505
Forks	1,831
Last release	rea-agents-5.0.0 (2026-10-07)
Languages	TypeScript 84.7%, JavaScript 9.1%, Java 2.1%
Contributors	18
AI contributor accounts (Claude / Codex / Cursor / Copilot)	Claude: 72, Cursor: 17
AI co-authored commits	Claude: 80, Cursor: 17

Main contributor

Field	Value
Name	morluto
GitHub	@morluto
Commits to this repo	630 (top contributor)
Bio	MTS @ _____
Website	https://x.com/morluto
Followers	437
Public repos	162
On GitHub since	2020-12-22

--------------------------------------------------------

Discover the best GitHub repos without the endless scrolling.

Get top repos organized by category (AI, DevOps, Databases & more), plus key takeaways from developers' most useful comments - all in one free, twice-weekly newsletter.

👉 Get the best repos & community insights - Free

checkout this amazing MCP related GitHub repo - 100% open source and free.

codebase-memory-mcp is a local code intelligence MCP server. It indexes a typical repo in milliseconds, the Linux kernel in about 3 minutes, and answers structural queries in under a millisecond, with no API key.

Post link / GitHub repo in comments below.

Features

• 161 languages via tree-sitter, plus Hybrid LSP
• 15 MCP tools: search, trace, impact, Cypher, dead code, ADRs
• About 120x fewer tokens than file-by-file grep
• 3D graph UI at localhost:9749
• 100% local SQLite, single static binary

Install

curl -fsSL https://raw.githubusercontent.com/DeusData/codebase-memory-mcp/main/install.sh | bash

Restart your coding agent and say "Index this project".

Clearly, the graph shows that the uptime for this model is abysmal, but OpenRouter reports ~99% uptime. How is it possible?

I've tried to use it for coding, but the results have been pretty poor. Still, on Openrouter it's used a lot: 2.58 Trillion tokens in total. IMO though, Hy3-free was a much better model. I miss that one.

What GPU would be needed to run the model deepseek v4 flash? I need at least 20 tokens/s. And how to set this up using vast.ai ?

Baidu quietly dropped something huge: Unlimited‑OCR, a long‑context OCR model that can read entire multi‑page documents in one shot — not page‑by‑page like traditional OCR tools.

This thing processes 40+ pages at once, keeps reading order, preserves tables across page boundaries, handles formulas, and outputs clean structured Markdown.

And it runs 100% locally.

🔍 Why this matters

Most OCR systems (even the big cloud ones) break documents into pages. That means:

• Tables split across pages get mangled
• Reading order gets lost
• Cross‑page context disappears
• You pay per page for cloud OCR (Textract, Cloud Vision, Azure Document Intelligence)

Unlimited‑OCR flips the script:

• 3B parameters, but only 500M active during inference
• 32K context window
• Multilingual
• Free & local
• 93% accuracy (+6 points over baseline)
• Error rate stays below 0.11 even past 40 pages
• 2.12M downloads on Hugging Face last month
• 14.6K GitHub stars

This is basically “OCR that reads like a human.”

🔗 Links

• GitHub: https://github.com/baidu/Unlimited-OCR
• Hugging Face: https://huggingface.co/baidu/Unlimited-OCR
• Technical report on arXiv: https://arxiv.org/abs/2606.23050

🧠 Use cases that suddenly become easy

• Parsing entire research papers
• Extracting tables from multi‑page financial reports
• Digitizing books
• Processing long PDFs without losing structure
• Building local document‑understanding pipelines without cloud costs

💬 What do you think?

Is this the beginning of long‑context document AI becoming mainstream?
Are we about to see local OCR pipelines replace cloud OCR entirely?

We just shipped Torvian Chatbot v0.8.0, and the headline feature is a brand-new Built-in Worker Tools system that brings first-class, sandboxed tool execution to your self-hosted AI workspace.

For anyone running their own AI stack, the hard part has never been the model — it's safely letting an agent do things: read files, run commands, fetch the web. v0.8.0 tackles exactly that, with three ideas we think matter a lot for self-hosters:

🔒 Workspace isolation by default. Built-in tools (file ops, run_command, web access) execute inside a configured workspace root. Filesystem escapes are blocked, and web tools reject localhost, loopback, and private-network addresses. Your agent can act without being handed the keys to your whole machine.

✍️ Signed authorization (E2EA). Tool executions are authorized through a signed payload validated by the worker's verification service. You register trusted signers (clients) via CLI or env vars, so only approved devices can trigger tool runs. No more trusting the network.

🧰 Full management UI + human-in-the-loop. A new Settings tab lets you enable/disable tools per worker, set per-tool auto-approve / auto-deny preferences, edit schemas, and reset to catalog defaults. Tool calls still require approval by default — you stay in control of every agentic action.

Under the hood this replaces the old generic tool paths with a shared BuiltInToolCatalog so advertised schemas stay in sync between server (seeding/LLM discovery) and worker (execution), plus a Flyway migration and a server→worker dispatch service with typed events and timeouts.

Torvian remains fully self-hosted: bring your own LLM providers (OpenAI-compatible, Ollama), keep your data on your hardware, and chat from Desktop, Web (WASM), or Android.

👉 Downloads & full notes: https://github.com/Torvian-eu/chatbot/releases 👉 Live demo: https://chatbot.torvian.eu/demo.html

We'd love to hear how you're using (or planning to use) worker tools — what would make this more useful for your setup? Questions and feedback welcome below!

We’ve released Torvian Chatbot v0.7.0, and the main focus of this update is message search across the entire workspace.

For a self-hosted AI setup, search matters a lot. As conversations grow longer and branch into threads, useful answers can get buried fast. With v0.7.0, Torvian Chatbot makes it much easier to find past context without giving up the core things that matter most: self-hosting, data control, and human-in-the-loop workflows.

Here’s what’s new:

• Search inside a chat session You can now search within the current conversation using a top-bar search UI, keyboard shortcut, result navigation, and highlighting. That means less scrolling through long chats and faster recovery of important prompts, answers, or tool outputs.

• Search across all chat sessions v0.7.0 adds a new cross-session search flow, so you can find messages across your workspace from one place. This is especially useful if you use Torvian as an ongoing AI workspace rather than a disposable chat tool.

• Thread-aware search Search can respect the currently visible branch or expand to all threads. That’s important in Torvian because conversations can branch, and users often want either a focused search or a broader one depending on the task.

• Navigate without losing context When jumping to search results in another thread/session, Torvian can bring you back to the previous conversation context. That rollback behavior makes search practical for real work instead of feeling disruptive.

Under the hood, this release also includes server-side search endpoints, ownership-scoped results, better state handling, and protections against stale overlapping responses.

If you’re looking for a self-hosted AI workspace with control over your models, your data, and your tool execution, this release makes day-to-day usage much smoother.

We’d love feedback on the search experience, especially from people using long-running chats, threaded workflows, or MCP tools. Questions, ideas, and edge cases are very welcome.

CTA: Learn more and download v0.7.0 here: https://github.com/Torvian-eu/chatbot/releases | Check out the demo: https://chatbot.torvian.eu/demo.html

Hashtags: #OpenSource #SelfHosted #AI #Search #Chatbot