AI Agent GitHub Digest — 2026-09-20

TL;DRaffaan-m/ECC rode agent-harness optimization to 260k+ stars in eight months, though a growth rate that steep deserves skepticism; cactus-compute/needle trades chat ability for tool-calling precision in an 8-29MB model; Graphify-Labs/graphify builds knowledge graphs with local AST parsing instead of a vector store; tinyhumansai/openhuman makes 'getting to know the user' the core of its agent memory; IvanMurzak/Godot-MCP lets agents drive the Godot editor directly; Claude Code v2.1.277 adds AGENTS.md support

#ai-agent #github #open-source #daily #agent-harness #mcp-server #agent-memory

Series: AI Agent GitHub Digest (36 / 36)

AI Agent GitHub Digest — 2026-09-19 AI Agent GitHub Digest — 2026-09-21

Today's Highlights

Today's trending repos split into two directions. One arms the agent itself: affaan-m/ECC optimizes harness performance, Graphify-Labs/graphify turns a codebase into a knowledge graph, and tinyhumansai/openhuman makes "get to know you first" its memory layer's core pitch. The other pushes the agent's reach into places it couldn't go before: cactus-compute/needle squeezes a tool-calling model into 8-29MB for phones and robots, and IvanMurzak/Godot-MCP lets agents drive a game engine's editor directly. Rather than "one more framework," this looks like the agent ecosystem expanding deeper and wider at the same time.

affaan-m/ECC ⭐ 262,767

GitHub　·　JavaScript　·　MIT

What it is: A performance-optimization plugin system for coding agents like Claude Code, Codex, and Cursor, bundling skills, "instincts" (behavioral tendencies), memory, and security into a pluggable harness layer.
Why it matters: Created in January 2026, it's already at 262k+ stars and 39k forks eight months later — one of the fastest-growing projects in the "add a harness to a general-purpose agent" category. But that speed is worth flagging: the repo carries a "GitHub Trending Repository of the Day" badge and ships its README in 12 languages simultaneously, and growth at this scale usually comes with some deliberate push for visibility, so the star count itself is worth taking with a grain of salt. The design — layering skills, instincts, and memory — does hit a real gap: general-purpose agents genuinely lack a harness.
Tech stack: JavaScript + integration layer for Claude Code / Codex / Opencode / Cursor
Getting started: Easy — install via the official GitHub App flow described in the README to hook it into an existing coding agent.

cactus-compute/needle ⭐ 11,533

GitHub　·　Python　·　Apache-2.0

What it is: An automation foundation model for phones, wearables, robots, cars, and microcontrollers — the whole model is an 8-29MB, 2-bit quantized binary that handles tool calling, structured extraction, and text embeddings.
Why it matters: The thinking runs opposite to mainstream "shrink the big model" work: instead of starting with a general chat model and compressing it, Needle decides upfront that the model does exactly three things — tool calling, extraction, embeddings — and trades away conversational ability for precision at those three. The team's own benchmarks claim exact-match tool-calling accuracy beating models 10x its size.
Tech stack: A Laddered Simple Attention Network (a Monarch Hadamard MLP in place of the FFN, GQA attention, and engram n-gram memory)
Getting started: Easy — pip install cactus-needle, or try the interactive browser demo.

Graphify-Labs/graphify ⭐ 119,589

GitHub　·　Python　·　Apache-2.0

What it is: Turns an entire codebase — docs, SQL schemas, configs, and PDFs included — into a queryable knowledge graph, shipped as a /graphify skill for Claude Code, Cursor, Codex, and Gemini CLI.
Why it matters: It takes the local, deterministic AST-parsing route instead of a vector store, which is the real difference from typical RAG tools — every edge in the graph traces back to a source, rather than being an embedding-similarity guess. The team is a YC S26 company, and GitHub Trendshift has it on the trending list too.
Tech stack: Python + tree-sitter AST parsing + Leiden community detection
Getting started: Easy — install from PyPI and run a skill command; the full platform's early-access v1 is still in testing.

tinyhumansai/openhuman ⭐ 39,887

GitHub　·　Rust　·　GPL-3.0

What it is: An open-source agent harness built around local-first memory — the pitch is spending time upfront so the agent learns how you work, instead of starting cold every conversation.
Why it matters: It frames itself against Karpathy's LLM Knowledgebase idea, and the difference from similar "memory across sessions" tools (like claude-mem) is that "getting to know the user" is the product's core, not a bolt-on feature. It ships under GPL-3.0, stricter than the MIT/Apache licenses most comparable projects use — worth checking before commercial use.
Tech stack: Rust + Tauri desktop app
Getting started: Moderate — it's a desktop app you install and run locally, not a pure CLI or MCP server.

IvanMurzak/Godot-MCP ⭐ 247

GitHub　·　C#　·　Apache-2.0

What it is: Connects the Godot game engine to MCP, letting agents like Claude, Cursor, and Copilot operate the Godot Editor directly — creating nodes, editing scenes, managing resources and scripts, capturing debug screenshots.
Why it matters: It shares the same MCP/reflection stack as the same author's Unity-MCP, via the ReflectorNet NuGet package rather than a separate rewrite, and exposes 42 built-in tools across 12 tool families. Complete solutions for "writing games with an agent" are still rare, and this one covers editor operations more thoroughly than most.
Tech stack: C# editor addon + ReflectorNet reflection framework + a self-hosted or cloud (ai-game.dev) MCP server
Getting started: Moderate — install the addon in the Godot Editor and configure the MCP connection (self-hosted or to the cloud backend).

Notable Releases

Claude Code v2.1.277 / v2.1.278

Release Notes

Key changes: v2.1.277 added AGENTS.md support — when a project has no CLAUDE.md, Claude Code reads AGENTS.md instead (switchable under "Project instructions" in /config; not yet available on Bedrock, Vertex, or Foundry). v2.1.278 changed the default auto-mode classifier for Claude API and Enterprise users, and for Bedrock, Vertex, Foundry, and gateway setups, to run server-side, which no longer bills for classifier overhead (CLAUDE_CODE_AUTO_MODE_SERVER=0 opts back out on Bedrock/Vertex/Foundry/gateways), and warns on any billed fallback.
Breaking changes: None — pure feature additions plus a large batch of bug fixes, including several crash fixes and a fix for claude -p / Agent SDK sessions hanging.
What it means for you: If your project already uses CLAUDE.md, this upgrade doesn't change anything. If your team follows the AGENTS.md convention, you no longer need to convert it. Teams running Bedrock/Vertex/Foundry or a self-hosted gateway who care about auto-mode billing should take a look at the CLAUDE_CODE_AUTO_MODE_SERVER toggle.

Today's Takeaway

I used to gauge "agent ecosystem growth" mostly by how fast new frameworks appeared. Today's picks — needle's edge devices, Godot-MCP's game engine — are a reminder that the more interesting signal might be which places the agent's reach is extending into. Once the framework layer starts to saturate, the real differentiation shows up in deep, vertical integrations, not another general-purpose orchestration layer.