AI Open Source Radar — August 6, 2026

Selection: AI repositories with the steepest recent (yesterday/today) growth. Ranked by Trendshift daily momentum and cross-checked against findarepo’s measured 7-day growth (both refreshed Aug 5). Large-but-stalled repos and previously featured repos whose growth has cooled are excluded. Data window: 21:08 UTC Aug 4 – 21:08 UTC Aug 5.

firecrawl/anydoc — #1 daily momentum (new breakout)

Trendshift daily #1 · ~1.4k ⭐ · github.com/firecrawl/anydoc

A Rust-based document parsing engine that Firecrawl open-sourced on August 4, converting Word, PowerPoint, Excel, OpenDocument, RTF, EPUB, CSV and PDF — 13 formats in all — into clean Markdown. Each format is parsed into a shared document model and then rendered through a single Markdown serializer, so escaping, tables, heading anchors and footnotes behave identically whether the input is a 2003 .doc or a .pptx from yesterday. It runs as pure Rust with no ML models or external services, hitting a median conversion time under 5ms per document, and handles text-based PDFs locally through the sibling pdf-inspector project without any OCR service. Node.js bindings run on the libuv thread pool so they never block the event loop, Python releases the GIL, and the whole thing is 100% open source — it powers Firecrawl’s own /parse endpoint. The appeal driving its surge: agents can turn documents into Markdown locally at extreme speed and feed them straight into RAG and ingestion pipelines.

Practical angle: convert large volumes of internal documents (contracts, decks, spreadsheets) to Markdown locally — no API keys, no cloud round-trips — cutting RAG-indexing preprocessing cost and latency.

Tags: #DocumentParsing #RustLang #MarkdownConversion #RAGPreprocessing #NodePythonBindings

TencentCloud/TencentDB-Agent-Memory — #2 daily momentum (2nd day surging)

Trendshift daily #2 · findarepo +5.2k / 7d · ~15k ⭐ · github.com/TencentCloud/TencentDB-Agent-Memory

A team-level memory hub for AI agents that turns conversations, docs and code into four reusable, governed memory assets shared and equipped across agents and frameworks:

Code-Graph — import a codebase and it auto-indexes symbols, files, call relationships and impact paths.
LLM-Wiki — import documents and it auto-generates structured pages with a link graph.
Skill and Chat Memory — auto-extracted from past agent conversation sessions as reusable assets.

The architecture rests on two pillars — memory layering and symbolic memory — building a semantic pyramid (L0 raw conversation → L1 atomic facts → L2 scenario blocks → L3 persona) so agents don’t merely “remember more” but “reason better.” It is explicitly designed to solve three problems: what is worth keeping, who is allowed to use it, and how to retrieve less while retrieving the right things next time. Its growth held and accelerated versus Aug 5, keeping it near the top for a second straight day.

Practical angle: use it as a central memory layer where multiple in-house agents share project knowledge, code graphs and work rules — cutting duplicate retrieval and token waste while making answers more consistent across a team.

Tags: #AgentMemory #CodeGraph #SymbolicMemory #KnowledgeGraph #TeamMemoryHub

trycompai/crm — #3 daily momentum (2nd day surging)

Trendshift daily #3 · ~5.4k ⭐ (tracking started) · github.com/trycompai/crm

An open-source, agentic-first CRM: instead of bolting AI onto a CRM as a feature, the CRM itself is where the agent keeps its notes. The agent runs on its own deployment, its own schedule and its own work queue — deciding what to look at next, booking its own follow-ups and spending a defined research budget. It is background-first rather than request/response, so it keeps working even after you close the browser. A guiding principle is that nothing about a person is guessed: tools report only what they observed, and a ledger prices the evidence. It is single-tenant and internal by design, using Google sign-in plus an allow-list as its authorization model, and every external source is optional so it can run with no API keys at all. Its increase has actually grown since the Aug 5 feature, extending a second straight day of surging.

Practical angle: a sales/CS team can self-host it — with no API keys — as a “CRM that works on its own,” handling lead research and follow-up scheduling continuously in the background.

Tags: #AgenticCRM #BackgroundAgents #SelfHosted #WorkQueue #EvidenceLedger

kirodotdev/KiroCrew — #6 daily momentum (new breakout)

Trendshift daily #6 · github.com/kirodotdev/KiroCrew

A persistent development workspace that self-improves and keeps working between conversations rather than ending when a chat closes. Sessions, memory, schedules and task checkpoints survive Gateway restarts, and scheduled or reactive work continues with nobody at the terminal. The core component, the Gateway, manages the lifecycle of agent sessions, handles incoming messaging events, and orchestrates background work through a GatewayOrchestrator. Self-learning is built in: corrections and task failures become durable lessons, preferences and project context carry into new sessions, and repeated patterns can be promoted into reusable skills. Its memory is organized into preferences, active project context, decaying history summaries and durable lessons; it orchestrates agents over the open Agent Client Protocol (ACP), and runs wherever you control the hardware — a Mac, a local container, or a remote machine.

Practical angle: set it up as a resident development co-worker that keeps triaging issues, updating docs and running recurring tasks through nights and weekends without supervision.

Tags: #PersistentAgents #AgentClientProtocol #SelfHosted #TaskScheduling #SelfImproving

cloudflare/computer — #10 daily momentum (new breakout)

Trendshift daily #10 · early preview (Aug 3–4) · github.com/cloudflare/computer

An open-source library (early preview) that gives each AI agent its own virtual working computer, complete with a filesystem and command-execution environment. Its virtual filesystem is backed by SQLite and can be populated from cloud storage, source control or any files you choose; the agent can read, write and edit files, run shell commands, and interact with Git repositories. That filesystem lives inside a Durable Object, which holds the authoritative state in SQLite and exposes a single pluggable execution surface via workspace.runtime. The runtime dynamically orchestrates between fast, efficient isolates and full Linux containers, so the agent always runs on the right compute primitive for the task. It is still an early preview — the API and design may change and it is not recommended for production — but the framing (“your agent needs a computer, not a container”) plus Cloudflare’s weight are driving the attention.

Practical angle: give coding and workflow-automation agents an isolated filesystem and shell, scaling up to full Linux only when a task needs it — a cost-efficient runtime without keeping containers always-on.

Tags: #AgentRuntime #DurableObjects #VirtualFilesystem #SQLiteFS #IsolatesVsContainers

Exclusions (for transparency)

reverse-skill (daily #9, +9.6k/7d): top-tier growth, but a dual-use reverse-engineering / penetration-testing security repo — held out as before out of caution.
kimi-k3-in-c, genoffice, yc-software/qm (featured Aug 5 / Aug 3): dropped from the top of the daily ranking / growth cooling — excluded.
bojieli/ai-agent-book (+6.7k/7d): a book-style educational repo, not a tool — excluded.
OmniRoute, stablyai/orca, airllm, pi, graphify, ponytail, strix, Agent-Reach, DeepSeek-Reasonix: large / mature / always-near-the-top or echoes of past spikes — excluded.