AI Open Source Radar — August 7, 2026

Only repos that are surging right now. Primary signal: recent star growth (today’s stars and daily momentum rank). Data window (UTC): 2026-08-05 21:07 to 2026-08-06 21:07. Source: Trendshift daily (freshness verified). Cumulative stars are a secondary reference. Local publication date 2026-08-07.


PrimeIntellect-ai/prime-agent — +1,800 ⭐ today (daily #1 · new)

Cumulative ~1.8k · github.com/PrimeIntellect-ai/prime-agent

Prime Agent is a self-improving RLM agent harness for coding workflows and even long-running autonomous tasks. Its core is a persistent Python control environment (persistent IPython) plus durable harness state, so useful working context and reusable operating patterns outlive a single chat window. Everything is programmatic: file operations, shell commands, tool use, sub-agents, and context management all happen through code, and rlm(...) spawns real child agents for parallel or background work that can discover one another and exchange messages to steer active work. Key mechanics:

  • /refine reviews the current trajectory and applies small, evidence-backed updates to supplemental harness state — never rewriting the immutable base system prompt, with snapshots for rollback.
  • /goal keeps an objective and its progress active across turns until completed, paused, or cleared.
  • Daemon-backed continuity: sessions, IPython state, schedules, and sub-agents keep running when the terminal detaches and can be reattached later.
  • Executable skills: skills are importable Python packages, and a built-in skill creator turns recurring workflows into project or personal skills.

Practical angle: a fit for teams that need long-running agents whose state survives detached, overnight or weekend runs — refactors, large migrations, continuous monitoring pipelines.

Tags: #RLMHarness #SelfImprovingAgent #PersistentIPython #MultiAgentMessaging #LongRunningAgents #ProgrammaticToolCalling


cloudflare/cloudflare-os — +1,600 ⭐ today (daily #2 · new)

Cumulative ~1.6k · github.com/cloudflare/cloudflare-os

Cloudflare OS is an agent workspace — built on Cloudflare Workers — for creating documents, building apps, and running agents with your company’s context and systems. It makes heavy use of Durable Objects, Dynamic Workers, and Facets: every workspace is its own Durable Object, every Gadget runs in a Dynamic Worker Facet, and Gatekeepers install facets to manage access to remote services. Its distinguishing premise is security architecture: agents receive scoped capability objects representing access to specific resources under specific policies — not raw API keys, OAuth tokens, or database credentials. A Gatekeeper (a Cloudflare Worker) sits between Cloudflare OS and any external service, holds that system’s credential, and the agent never sees it. Released under Apache 2.0, the intent is that companies copy and customize it into “Your Company OS” rather than adopt it as a SaaS.

Practical angle: well suited to security-sensitive organizations building an internal AI workspace that connects to in-house SaaS and databases under least privilege — without ever handing credentials to the agent.

Tags: #CloudflareWorkers #DurableObjects #ScopedCapabilities #CredentialIsolation #GatekeeperPattern #EnterpriseAgentWorkspace


cloudflare/computer — +1,400 ⭐ today (daily #3 · 2 days running, accelerating)

Surging · github.com/cloudflare/computer

@cloudflare/computer gives each agent its own virtual computer. It combines a SQLite-based virtual filesystem with a Durable Object and dynamically orchestrates between a lightweight isolate and a full Linux container depending on the workload, providing an isolated, persistent environment where an agent can keep files and run code. Because state lives in the Durable Object, files and working context survive a dropped session, and the heavier container is only spun up when needed to balance cost and performance. First featured yesterday (08-06) as a new entry, today it jumped from around #10 to #3 with growth actually increasing — accelerating — so it earns a second day on the radar. Alongside sister project cloudflare/cloudflare-os, it anchors Cloudflare’s emerging “agent infrastructure” lineup.

Practical angle: drop-in for code-execution or sandboxed-tool services that need per-agent isolated execution and file environments.

Tags: #PerAgentSandbox #DurableObjects #SQLiteVFS #IsolateToContainer #AgentRuntime #StatefulExecution


firecrawl/anydoc — +1,300 ⭐ today (daily #4 · 2 days running, still surging)

Surging · github.com/firecrawl/anydoc

anydoc is a Rust document-parsing engine that converts 13 formats — Word, PowerPoint, Excel, OpenDocument, RTF, EPUB, CSV, PDF and more — into clean Markdown. Written in Rust, it targets very fast processing (around <5ms per document), and handles PDFs locally through its sister project pdf-inspector. It ships Node.js and Python bindings so it slots easily into existing pipelines, making it especially useful for normalizing varied office documents into consistent Markdown during RAG, document indexing, and LLM input preprocessing. It debuted at daily #1 yesterday (08-06) and today holds a strong ~+1.3k gain at #4 — only edged down in rank by newer breakout repos while its own momentum holds — so it stays on for a second day.

Practical angle: deploy it as a preprocessor to quickly standardize mixed-format files into Markdown before feeding a RAG or search index.

Tags: #RustLang #DocumentToMarkdown #RAGPreprocessing #PDFParsing #OfficeFormats #NodePythonBindings


Binaryify/open-kimi-ppt-skill — +456 ⭐ today (daily #7 · new)

Surging · github.com/Binaryify/open-kimi-ppt-skill

open-kimi-ppt-skill is an unofficial Kimi Slides skill that helps AI agents generate genuinely editable slide decks. Its core is the PPTD intermediate format, which describes themes, layouts, and elements in YAML — more stable than writing OOXML/pptxgenjs directly and more locally editable than image-based PPT generation. How it works:

  • Dual output: an editable PPTD project for continued iteration plus a ready-to-use PPTX with embedded fonts and fade-in/fade-out transitions.
  • True editability: after export, text boxes and shapes stay editable in PowerPoint or WPS — unlike image-based PPTs usable only as posters.
  • Local visual editor: a browser-based editor for previewing, fine-tuning, configuring transitions, and re-exporting manually without re-running the agent.
  • Compatibility: works with any agent supporting the SKILL.md spec — Codex, Claude Code, Cursor, and more.

(Note: this is an unofficial project based on reverse-engineering Kimi Slides, not an official Kimi/Moonshot AI project.)

Practical angle: good for sales/consulting deck automation where a human must finish an agent’s draft — the output is a real, editable PPTX ready to hand off.

Tags: #KimiSlides #EditablePPTX #PPTDFormat #SkillMdSpec #SlideGeneration #BrowserEditor


Deliberately excluded (not surging / decelerating)

  • trycompai/crm: featured 08-06 on a ~+4k one-day explosion, but today’s gain collapsed to +638 and its daily rank slid from #3 to #16 (momentum broken) — excluded.
  • TencentCloud/TencentDB-Agent-Memory: featured two days (08-05, 08-06), but now ~16k cumulative, with rank falling #2 to #18 and growth slowing — excluded.
  • Large / mature / perennial names: crawl4ai, earendil-works/pi, DietrichGebert/ponytail, esengine/DeepSeek-Reasonix, CherryHQ/cherry-studio, donnemartin/system-design-primer — excluded.
  • In the momentum pack but edged out of the top 5: huangruiteng/loopx, get-bb/bb, KKKKhazix/human-writing — watch-only.

Leave a comment