AI Open Source Radar — August 12, 2026

Ranked by recent daily growth on GitHub Trending (2026-08-11). A streak log damps repeat appearances — repos whose momentum has stalled are dropped; only those holding or accelerating return.

PrimeIntellect-ai/prime-agent — +2,642 ⭐ today (~7k total)

Prime Agent is a self-improving, open-source agent harness for coding and long-running autonomous work. Its core is an RLM (Recursive Language Model) that treats context as variables and treats tools and subagents as function calls inside a persistent IPython REPL. File edits, shell commands, tool use, subagents and context management all happen through code, and an rlm(...) call spawns real child agents for parallel or background work, returning their results programmatically. A Continual Harness stores supplemental prompts, memories, skill descriptions and reusable subagent specs as durable state, and /refine applies small, evidence-backed updates to that state without ever rewriting the immutable base system prompt (snapshots support rollback). Released on August 5 and No. 1 on GitHub Trending by August 7, it drew repetition scrutiny before re-accelerating today to +2,642 — above its debut spike.

Recursive Language Model: context as variables; tools and subagents as function calls
Persistent IPython REPL as the built-in model tool
Continual Harness: durable memories, skills and reusable subagent specs
/refine: evidence-backed self-updates with snapshot rollback

Where it fits: parallelize long-running refactors or research pipelines across subagents. Repo: github.com/PrimeIntellect-ai/prime-agent

Tags: #RLM #CodingAgent #SubagentOrchestration #SelfImproving #PersistentREPL #ContinualLearning

diegosouzapw/OmniRoute — +975 ⭐ today (new / surging)

OmniRoute is a free, MIT-licensed AI gateway that reaches 290-plus providers (90-plus of them free) and 500-plus models — Kimi, Claude, GPT, Gemini, GLM, DeepSeek, MiniMax and more — through a single endpoint. Two mechanisms define it: quota-aware auto-fallback silently slides to the next model when a quota runs out, a provider fails, or costs spike; and RTK+Caveman compression cuts tokens by 15–95% (with a default inflation guard, multi-language Caveman packs, and RTK filters for Gradle and .NET). It plugs into existing coding tools — Claude Code, Codex, Cursor, OpenCode, Cline and Copilot — and ships MCP/A2A support plus desktop and PWA clients. Shaped by 500-plus contributors, it is surging on demand for stacking free tiers and driving inference costs down.

One endpoint, 290+ providers and 500+ models
Quota-aware auto-fallback on quota exhaustion, provider failure or cost spikes
RTK+Caveman compression saves 15–95% tokens
Works with Claude Code, Codex, Cursor, Cline, Copilot; MCP/A2A, desktop and PWA

Where it fits: consolidate multiple LLM-vendor accounts behind one endpoint to reduce cost and availability risk. Repo: github.com/diegosouzapw/OmniRoute

Tags: #LLMGateway #ModelRouting #TokenCompression #QuotaFallback #MultiProvider #ClaudeCodeCompatible

semantica-agi/semantica — +970 ⭐ today (~4.7k total)

Semantica is graph-native infrastructure for explainable, auditable AI — an open-source Python library under MIT. The point is that it stores meaning, not embeddings: what an agent knows lives in a structured, queryable context graph; every decision is tracked as a first-class object; and every fact links back to its source with full provenance (W3C PROV-O compliant). It ships reasoning engines — forward chaining, Rete, deductive, abductive and SPARQL — and runs as a deterministic layer that needs no LLM for graph construction, reasoning or provenance. Because it sits underneath your LLM, vector store and agent framework, it drops into LangGraph, CrewAI or LlamaIndex, and it is climbing today on rising demand for accountable, regulation-ready AI.

Structured, queryable context graphs instead of opaque embeddings
Decisions as first-class, trackable objects
Full provenance back to source (W3C PROV-O)
Reasoning engines: forward chaining, Rete, deductive, abductive, SPARQL — no LLM required

Where it fits: add a provenance layer to agents at finance or healthcare customers that need decision rationale and audit trails. Repo: github.com/semantica-agi/semantica

Tags: #KnowledgeGraph #Provenance #DecisionIntelligence #ExplainableAI #ProvO #DeterministicReasoning

vitali87/code-graph-rag — +682 ⭐ today (~1.7k total, day 2)

Code-Graph-RAG lets you query, understand and edit a mixed-language monorepo in plain English. It parses the codebase with Tree-sitter, builds a structural knowledge graph in Memgraph, and translates natural language into Cypher to query and modify that graph. A single unified schema covers Python, TypeScript, JavaScript, Rust, Go, Java, C/C++, C#, PHP, Lua and Dart (Scala in progress; Ruby structurally supported via an ast-grep tier). Recent additions — an MCP server for Claude Code and intent-based semantic code search — shrink the context you feed AI coding tools down to graph queries. It grew from +328 yesterday to +682 today, extending a two-day climb.

Tree-sitter parsing to a Memgraph knowledge graph
Natural language to Cypher for querying and editing
One schema across 12+ languages
MCP server for Claude Code plus semantic code search

Where it fits: tame context blow-up for AI coding tools on large monorepos, improving both accuracy and cost. Repo: github.com/vitali87/code-graph-rag

Tags: #CodeRAG #TreeSitter #Memgraph #Monorepo #MCP #CypherQuery

danielmiessler/LifeOS — +315 ⭐ today (new)

LifeOS is a general-purpose AI harness for life and work that calls itself an “intent engineering” platform. It captures who you are and what you ultimately want, then conveys that to your AI on every task, organized around one central idea: moving from your Current State to your Ideal State. A memory system — hot-layer memory plus a typed Knowledge Archive (People, Companies, Ideas, Research), learnings and work history — consolidates across sessions so each one starts smarter than the last. Named subsystems include Synapse (input router), Conduit and Feed (internal and external senses), Atlas (a live asset graph) and Ledger (change tracking). It is AI-native — you install it into a coding harness such as Claude Code with a one-liner — and it is climbing as the latest project from fabric creator Daniel Miessler, riding the “personal AI infrastructure” wave.

Intent engineering: capture ultimate goals, convey them on every task
Current State to Ideal State as the organizing concept
Hot-layer memory plus a typed Knowledge Archive, consolidated across sessions
Subsystems: Synapse, Conduit/Feed, Atlas, Ledger

Where it fits: a template for structuring a person’s or team’s goals and context so agents receive consistent “intent.” Repo: github.com/danielmiessler/LifeOS

Tags: #IntentEngineering #PersonalAI #MemorySystem #AIHarness #KnowledgeArchive #ContextGraph

Excluded today: daily_stock_analysis (+4,391 yesterday → +731 today, momentum sharply down); large/always-on trending repos (ComfyUI, firecrawl, google/skills); and non-AI tools (3b1b/manim, brightdata/cli).