AI Open Source Radar — July 13, 2026

Selection criteria: newly surging repositories with the steepest recent (yesterday/today) star growth. Repos with large cumulative stars but slowing momentum (“rank-holders”), and repos featured in a previous edition whose growth has since cooled, are excluded. Cumulative stars are a reference only.

1. esengine/DeepSeek-Reasonix — +1,100 ⭐ today

Cumulative 26.7k ⭐ · Go · github.com/esengine/DeepSeek-Reasonix

DeepSeek-Reasonix is a DeepSeek-native AI coding agent that runs in your terminal, designed around “prefix-cache stability.” Its agent loop is append-only and aligned to DeepSeek’s byte-stable prefix cache, so even long sessions hold a 90%+ cache-hit rate and input-token cost collapses to roughly one-fifth. DeepSeek ships as the default preset, but any OpenAI-compatible endpoint is just a config entry rather than new code, and you can run two models — an executor and a planner — in separate cache-stable sessions. It is plugin-driven: external tools run as stdio JSON-RPC (MCP-compatible) subprocesses while built-in tools self-register at compile time, and it ships as a CGO-disabled single binary that cross-compiles to six targets in one command (its only dependency is a TOML parser). The pitch of slashing token costs for always-on agents has driven a surge of about +1,100 stars today — the steepest riser on the board.

Append-only loop aligned to DeepSeek’s byte-stable prefix cache (90%+ cache hits)
Input-token cost cut to ~1/5 on long sessions
Optional dual-model executor + planner setup
Plugin tools over stdio JSON-RPC (MCP-compatible)
CGO-free single binary, six cross-compile targets, TOML-only dependency

Real-world use: cut API costs by up to 80% for long-running coding agents through prefix-cache alignment.

Tags: #DeepSeek #PrefixCache #TerminalAgent #CodingAgent #TokenEfficiency #MCP

2. iOfficeAI/OfficeCLI — +6.4k ⭐ this week

Cumulative 15k ⭐ · C# · github.com/iOfficeAI/OfficeCLI

OfficeCLI is an open-source Office suite purpose-built for AI agents to read, edit, and automate Word, Excel, and PowerPoint files, and it runs as a single binary with no Microsoft Office installation required. A built-in HTML rendering engine reproduces .docx/.xlsx/.pptx as HTML or PNG with high fidelity so an agent can actually “see” document layout, and a formula-and-pivot engine auto-evaluates 350+ Excel functions on write — including dynamic arrays, financial and statistical functions, and native OOXML pivot tables. It auto-detects and installs into major coding agents such as Claude Code, Cursor, Windsurf, and GitHub Copilot, and distributes via Homebrew, Scoop, and npm. Its differentiator is handing an agent full control of Office documents in a single line of code. It surged +6.4k stars over the past week as a fresh entrant.

Single binary, no MS Office install needed
HTML rendering engine: .docx/.xlsx/.pptx → HTML or PNG
350+ Excel functions with dynamic arrays and native OOXML pivot tables
Auto-installs into Claude Code, Cursor, Windsurf, GitHub Copilot

Real-world use: wire report, financial-model, and slide generation directly into agent workflows to build an Office automation pipeline.

Tags: #OfficeAutomation #DOCX #XLSX #PPTX #OOXML #PivotTables #SingleBinary

3. Panniantong/Agent-Reach — +4.1k ⭐ this week

Cumulative 55k ⭐ · Python · github.com/Panniantong/Agent-Reach

Agent-Reach gives an AI agent “eyes” on the internet through a single CLI, letting it read and search 12+ platforms — Twitter/X, Reddit, YouTube, GitHub, Bilibili, XiaoHongShu and more — with zero API fees. It acts as a capability layer that handles tool selection, installation, health checks, and routing, so the agent can read tweets, search Reddit, and browse YouTube or GitHub without configuration. Web pages go through Jina Reader and semantic search runs on Exa, and by combining free, open-source tools with browser-based access the only optional cost is a $1/month proxy when a network blocks a platform. Installation is zero-config: paste one install.md line into your agent. Its appeal is bypassing the cost and complexity of paid social APIs, and it has held +4.1k stars over the past week while still climbing.

12+ sources: Twitter/X, Reddit, YouTube, GitHub, Bilibili, XiaoHongShu, LinkedIn, RSS, and more
Capability layer: selection, install, health checks, routing
Jina Reader for web pages, Exa for semantic search
Zero API fees; only optional cost is a $1/month proxy

Real-world use: let market- and sentiment-research agents collect real-time social data without paid APIs.

Tags: #WebScraping #SocialMediaSearch #AgentPerception #ZeroAPIFees #CLITool

4. calesthio/OpenMontage — +3.8k ⭐ this week

Cumulative 37k ⭐ · Python · github.com/calesthio/OpenMontage

OpenMontage bills itself as the world’s first open-source, agentic video production system, turning an AI coding assistant into a full video studio with 12 pipelines, 52 tools, and 500+ agent skills. Every production request must pass through the pipeline system — no exceptions — and the AI agent reads pipeline manifests, stage-director skills, and meta skills, then drives tools to automate the same structured process a real production team follows. Web research is a first-class stage: the agent searches YouTube, Reddit, Hacker News, news, and academic sources, gathers data points, audience questions, and trending angles, and compiles a cited research brief so videos are grounded in real, current information. Its differentiator is end-to-end automation of ideation, research, and production with nothing more than a coding agent, and it surged +3.8k stars this week.

12 pipelines, 52 tools, 500+ agent skills
Mandatory pipeline system driven by manifests and stage-director skills
First-class web-research stage with cited briefs
Turns a coding assistant into an end-to-end video studio

Real-world use: automate YouTube and marketing videos from research through editing in one agent pipeline.

Tags: #VideoProduction #AgentSkills #ContentPipeline #CreativeAutomation #PipelineOrchestration

5. DeusData/codebase-memory-mcp — +3.8k ⭐ this week

Cumulative 30k ⭐ · C · github.com/DeusData/codebase-memory-mcp

codebase-memory-mcp is a high-performance code-intelligence MCP server that indexes a codebase into a persistent knowledge graph so AI coding agents can answer structural questions with roughly 120x fewer tokens. It parses 158 languages via tree-sitter AST analysis, adding hybrid LSP semantic type resolution for Python, TS/JS, PHP, C#, Go, C/C++, Java, Kotlin, and Rust to build a graph of functions, classes, call chains, HTTP routes, and cross-service links. It full-indexes an average repo in milliseconds and the Linux kernel (28M LOC, 75K files) in three minutes, answering structural queries in under 1ms. It contains no LLM — it is a pure structural-analysis backend that leaves the intelligence layer to MCP clients like Claude Code — and ships as a zero-dependency single static binary exposing 14 MCP tools. Running 100% locally with no telemetry, it was evaluated across 31 real repos at 83% answer quality, 10x fewer tokens, and 2.1x fewer tool calls versus file-by-file exploration.

tree-sitter AST across 158 languages + hybrid LSP type resolution
Average repo indexed in milliseconds; Linux kernel in 3 minutes; sub-1ms queries
No LLM — structural backend for MCP clients; 14 MCP tools
Zero-dependency single static binary, 100% local, no telemetry

Real-world use: a local index that sharply reduces an agent’s code-exploration tokens on large repos, lowering cost and latency.

Tags: #CodeIntelligence #KnowledgeGraph #TreeSitter #MCP #LocalFirst #StaticBinary

Note: yesterday’s picks — stablyai/orca, diegosouzapw/OmniRoute, and ogulcancelik/herdr — were dropped today as their growth stalled or cooled. Rank-holding giants such as obra/superpowers, mattpocock/skills, affaan-m/ECC, and OpenClaw are permanently excluded.