Ranking basis: growth over the last 24 hours (2026-08-17 22:13 UTC to 2026-08-18 22:13 UTC). Cumulative stars are context only. Sources: Trendshift daily and weekly boards plus individual repository pages.
Today’s shape: deepseek-ai/deepseek-harness, which dominated yesterday’s daily board at +13,800 stars, collapsed to +1,100 — a 92% drop. Its satellites fell in lockstep: deepseek-harness-desktop went from +2,200 to +219, awesome-dsh-plugin from +2,100 to +87, and needle dropped off the daily board entirely. All are excluded under the streak-decay rule. With the DSH wave receding, the board is quiet, and what filled the gap is a cluster of projects about folding agents into a team rather than running them solo.
yetone/cumora — +253 ⭐ today (+2,200 this week, #2 daily)
Cumulative stars: ~2,200 · TypeScript · MIT · https://github.com/yetone/cumora
Cumora is a cross-platform team chat that treats AI agents as colleagues rather than tools. Humans and agents share the same roster, DMs and group conversations, plus a common Kanban board and calendar. The backend is a stateless Node service built on Express and WebSocket, with Postgres as the single source of truth and Redis handling pub/sub fan-out and presence — so any number of instances behind a load balancer stay in sync. Each agent carries its own persona and memory, claims work directly, coordinates with other agents to avoid collisions, and can send and receive real email. There are two execution paths: cloud agents run in per-agent Kubernetes pods with tool-calling on the OpenAI Responses API, while BYOA (Bring Your Own Agent) pairs your own Mac or VPS with a daemon so the agent’s “brain” is your local Claude Code or Codex CLI on your existing subscription. Developer yetone — known for avante.nvim — open-sourced it yesterday and it went straight to #2 on the daily board as the idea spread on X: throw agents from several vendors into one group chat and hand out assignments.
Practical angle: a good fit for small teams that want to run coding agents from multiple vendors on their own subscriptions while keeping task assignment and progress in a single chat and Kanban view.
Tags: #AgentTeamChat #BringYourOwnAgent #KubernetesAgentRuntime #WebSocketPubSub #KanbanCoordination #AgentPersona
akitaonrails/ai-memory — +67 ⭐ today (+891 this week, ~40% weekly growth against its base)
Cumulative stars: ~2,100 · Rust · MIT · 22 contributors · https://github.com/akitaonrails/ai-memory
ai-memory is a single Rust binary that gives coding-agent CLIs long-term memory and lets work hand off between agents from different vendors. It covers Claude Code, OpenAI Codex, Cursor, Gemini CLI, Antigravity CLI, Grok Build CLI, OpenClaw, OpenCode, OMP and MCP-capable clients generally. The storage design is what sets it apart: instead of piling embeddings into a vector database, it compiles a git-versioned tree of Markdown pages on disk over time — what the author calls a Karpathy-style LLM wiki. Pages are versioned in place through supersession rather than overwritten, semantic concepts compound, and episodic logs decay with age. Because every prompt and tool call lands in ai-memory, opening the next session surfaces a handoff describing where you left off; switching harnesses starts or resumes the native session linked to the shared workstream, and an obsolete local session cannot overwrite newer cross-harness history.
- Run the server with
docker run - Wire up Claude Code with
ai-memory install-mcp - Wire up other agents with
ai-memory install-hooks
The project has been public since May, but it broke into GitHub Trending at #2 on August 17 and only caught fire this week.
Practical angle: where developers on the same team each use a different coding agent, this keeps project context in a human-readable Markdown wiki, cutting the cost of switching tools or handing work over.
Tags: #AgentMemory #CrossVendorHandoff #GitVersionedWiki #MCPServer #RustSingleBinary #SessionContinuity
ifixai-ai/iFixAi — +74 ⭐ today (+1,800 this week)
Cumulative stars: ~10,600 · Python · Apache 2.0 · https://github.com/ifixai-ai/iFixAi
iFixAi is an independent auditing tool for AI agents that aims to answer “is this agent actually doing its job?” in under 120 seconds — run either by a human or by the agent on itself. Its grading scheme groups 45 inspections into 16 categories, split into five core pillars and eleven premium ones, and returns a letter grade from A at 0.90 or above down to F below 0.60.
- Manipulation — weight 0.35
- Fabrication — weight 0.20
- Deception — weight 0.15
- Unpredictability — weight 0.15
- Opacity — weight 0.15
The sibling repository ifixai-ai/diagnostic is an open-source misalignment diagnostic that runs 32 tests across those same five axes. It is provider-agnostic — it points at OpenAI, Anthropic, Bedrock, Azure, Gemini and others without modification — and produces a content-addressed manifest for bit-identical replay. The pitch is that AI red teaming and operational assurance overlap, and iFixAi sits in that intersection: adversarial depth plus assurance discipline in one process, with evidence that survives review. It first trended at #2 among Python repos in early August and has held a weekly top position for two weeks, adding another +1,800 this week as organizations that have put agents into real workflows start shopping for audit and governance tooling.
Practical angle: run it as a pre-delivery gate before shipping an agent to a client, and attach the grade plus the replayable evidence manifest to the deliverable to smooth procurement review.
Tags: #AgentAuditing #AIRedTeaming #MisalignmentDiagnostics #ProviderAgnostic #ReplayableEvidence #AIGovernance
What Was Excluded, and Why
- deepseek-ai/deepseek-harness — +13,800 yesterday to +1,100 today (-92%). Streak-decay rule applied.
- anywhere-labs/deepseek-harness-desktop — +2,200 to +219 (-90%).
- awesome-dsh-plugin/awesome-dsh-plugin — +2,100 to +87 (-96%).
- cactus-compute/needle — dropped off the daily board.
- AlexsJones/llmfit (32.3k) and mukul975/Anthropic-Cybersecurity-Skills (28.6k) — near the top since February and March respectively, with daily gains of 70 and 64. Rank-holders, not risers.
- MoneyPrinterTurbo, Motrix, public-apis — large and mature.
- Promotional and non-AI entries on the board were skipped.
Three entries rather than five is the honest count for a quiet day — padding the list would defeat the point of a radar.
Daily AI Brief 
Leave a comment