Open Source
Open Source

AI Agents Want Group Chats Too: cumora by the avante Author Puts Claude Code and Codex on the Same Roster (1.4k Stars in One Day)

yetone/cumora (1,460 stars / 158 forks, MIT, TypeScript; created Aug 17, API same-day snapshot): the avante.nvim author's team chat for AI agents - agents as first-class teammates with personas, memory, atomic work claims, real email, and shared Kanban/calendar. Two brain paths: Cumora Cloud (per-agent K8s pods on the OpenAI Responses API) or BYOA (npx cumora agent computer - your local Claude Code/Codex on your own subscription; the server never sees keys). Anti-collision trio (seen-cursor freshness gate / atomic claims / small-brain triage), an llm_calls cost ledger, and a CI big-model guard. Five minutes local: Postgres + Redis + OPENAI_API_KEY. Five cautions include day-one maturity and cloud token burn.

Published August 18, 20268 min read
<!-- cumora-resource | open-source | AI Agents Want Group Chats Too: cumora by the avante Author Puts Claude Code and Codex on the Same Roster (1.4k Stars in One Day) -->

At 09:13 UTC on August 17, 2026, a new repository appeared on GitHub: yetone/cumora. Within a day: 1,460 stars, 158 forks (GitHub API snapshot, 2026-08-17). The author, yetone, is a Neovim-scene veteran - the creator of avante.nvim (the project that brought Cursor-style AI coding into Neovim). This time the new build isn't a coding plugin but something far more "absurd": team chat software for AI agents. Humans and agents share one roster, the same DMs, the same group conversations, the same Kanban board and calendar - the agent isn't a bot you @; it's a teammate with personas and memory that claims work.

Scope note: this article is based on the GitHub API snapshot and the official README (2026-08-17). Star counts are a same-day snapshot (they move fast; the repo is authoritative). The product is in very early, fast-moving shape - treat the docs as the source of truth.

1. What It Is: The Company Group Chat, for Agents

One line: cross-platform team chat where AI agents are first-class participants. This is not "stuff a bot into Slack"; it's a collaboration system designed with agents as full teammates:

  • Agents hold personas and memory - they aren't answer-machines that only respond when poked;
  • Agents claim work (atomic claims, no collisions) and share the Kanban board and calendar with humans;
  • Agents can send and receive real email (outbound via Resend, inbound via a Cloudflare Email Worker);
  • Humans and agents live on the same roster, in the same groups, with the same DMs.

The smartest design decision is the two "brain" paths:

PathWhere the brain livesFor whom
Cumora CloudOne managed K8s pod per agent, running a multi-hop tool-calling loop on the OpenAI Responses API (bash, files, browser, email, memory, skills)Out-of-the-box; people who don't want to run machines
BYOA (Bring Your Own Agent)You run npx cumora agent computer; the agent's brain becomes your local Claude Code or Codex CLI on your own subscriptionDevelopers with existing Claude/Codex subscriptions: the server never sees your provider keys

BYOA hits the pain point dead-on: Claude Code / Codex subscribers already pay monthly. cumora lets that subscription "stretch" into the brains of chat-room agents - no second per-token bill from a cloud agent.

2. The Architecture: A Small but Complete Distributed System

The README's architecture diagram is dense; worth unpacking:

  • Frontend: React 18 + Vite + TypeScript + Tailwind; one component set drives four shells - desktop (Electron), web (PWA), iOS and Android (Capacitor);
  • Backend: stateless Express + WebSocket service; Postgres is the single source of truth (Drizzle schema), Redis handles pub/sub fan-out and presence - any number of instances behind a load balancer stay in sync;
  • Agent runtime: cloud agents live in per-agent Kubernetes pods (orchestrated via kubectl from the server; a Go FUSE driver mounts the agent workspace into the pod); BYOA agents live on your machine. Every LLM call - cloud or BYOA - lands in one llm_calls cost ledger;
  • Edge: Cloudflare Workers for email-gate (inbound mail) and r2-gate (signed CDN).

The most interesting part is the anti-collision coordination (docs/COORDINATION.md): agents in the same room don't trample each other - the server arbitrates with a seen-cursor freshness gate (a stale reply is HELD and re-decides after seeing newer messages), atomic claims on real units of work, and a small-brain triage gate that shields the expensive big model. The repo ships real-LLM multi-agent coordination benchmarks (chain / counting / werewolf / kanban). CI even has a guard:big-brain check: only agent turns may use the big model - cost discipline enforced by pipeline, not by hope.

3. Install: Five Minutes Locally

Running locally needs Postgres + Redis + one OPENAI_API_KEY:

bash
createdb -h localhost cumora
export OPENAI_API_KEY=sk-...

npm install
npm run dev:all   # Vite renderer on :5180 + API server on :5181

Open http://localhost:5180 (PWA mode), or run npm run electron:dev for the desktop window. The schema is created idempotently on boot, and an empty database is seeded with a starter team: 6 agents, 3 humans, 9 conversations, and zero messages - everything that shows up in chat is produced live, not canned demo data. Every other env var (database, Redis, model choices, port) has a sane default; optional feature groups (OAuth, email, push, waitlist) soft-disable when unset.

For BYOA: on your own Mac/VPS run npx cumora agent computer and pair a local Claude Code / Codex as the agent's brain.

4. Five Cautions

  1. Born yesterday: repo created 2026-08-17 09:13 UTC, pushed several times the same day. Features, APIs, and docs are all in fast motion - don't expect a mature product.
  2. Cloud burns tokens: Cumora Cloud gives each agent a pod plus a multi-hop tool loop. Not cheap - luckily the llm_calls ledger keeps it visible. Watch it.
  3. BYOA's security boundary is yours to hold: the server never seeing your keys is true; but your local agent brain has bash. Before giving it group-chat reach, think through its tool-permission surface.
  4. Email is a double-edged sword: agents that send real email demo beautifully; configure sending domain, rate limits, and content review before anything production-adjacent.
  5. Don't take the demo as the benchmark: the repo ships coordination benchmarks (chain/counting/werewolf/kanban). Evaluate on your own tasks, not README screenshots.

5. Who It's For

  • Developers already subscribing to Claude Code / Codex: BYOA turns one subscription into a multi-agent team - the best value play right now;
  • Engineers building multi-agent systems: the coordination designs (freshness gate / atomic claims / small-brain triage / cost ledger / CI big-model guard) are a reference architecture worth stealing;
  • Small teams experimenting with "AI teammates": the seeded starter team (6 agents + 3 humans) works out of the box, saving a week versus wiring Slack bots + memory stores + schedulers from scratch.

Not for: enterprises wanting stable SLAs (give it a few months), or non-technical users who don't want to touch Postgres/Redis (wait for the cloud).

In a 2026 where agent frameworks are a red ocean, cumora flips the frame: don't build a stronger agent - build the office where agents work. Once everyone has a dozen agents, the problem shifts from "how to build agents" to "how to keep them from stepping on each other" - and that spot has almost no tenants yet.

FAQ

Q1: How is cumora fundamentally different from wiring an AI bot into Slack? A1: A bot is an appendage that answers when @-mentioned. cumora's agents are first-class: they hold personas and memory, proactively claim work, coordinate among themselves (freshness gating + atomic claims prevent collisions), exchange real email, and share the Kanban board and calendar. The collaboration substrate was rebuilt for human-agent parity.

Q2: In BYOA mode, can cumora's server see my Claude Code subscription? A2: No. The BYOA brain runs on your own machine (the npx cumora agent computer daemon); Claude Code / Codex execute locally against your subscription. The server only relays messages and never sees your provider keys.

Q3: Do I really need Postgres and Redis? A3: For local development, yes - Postgres is the single source of truth and Redis drives realtime fan-out and presence. The only hard-required env var is OPENAI_API_KEY (cloud mode); BYOA mode doesn't even need that.

Q4: Could the stars be inflated? A4: A day-one 1.4k is mostly author reputation cashing in (avante.nvim's creator) plus HN-style spread. Judge the repo on hard-to-fake engineering density: the architecture docs (BYOA/COORDINATION/SHIPPING), the coordination benchmarks, and the CI cost guard say more than the star count.

Q5: How does this relate to the agent frameworks this site has covered? A5: Frameworks (LangGraph, the harnesses) solve "how to build a single agent"; cumora solves "how a pile of agents and humans share one workspace". Engine versus office. For the infrastructure of routing and paying for models, see our Stripe-OpenRouter hotspot piece.


References

  • GitHub: yetone/cumora (1,460 stars / 158 forks, MIT, TypeScript; API snapshot 2026-08-17, repo created same day 09:13 UTC)
  • Official README: architecture diagram, BYOA path, coordination, local setup, repo layout (docs/BYOA.md, docs/COORDINATION.md, etc.)
  • Author: yetone (creator of avante.nvim; verified via GitHub profile)
  • npm: cumora (BYOA daemon, per README)
  • Related: Stripe-OpenRouter Hotspot, Agent Harness Comparison

Based on the official repository's public information (2026-08-17 snapshot); star counts are same-day API readings; features subject to official docs.

This article is AI-assisted and human-edited. Last updated: 2026-08-18

FAQ

How is cumora fundamentally different from wiring an AI bot into Slack?
A bot is an appendage that answers when @-mentioned. cumora's agents are first-class: they hold personas and memory, proactively claim work, coordinate among themselves (freshness gating + atomic claims prevent collisions), exchange real email, and share the Kanban board and calendar. The collaboration substrate was rebuilt for human-agent parity.
In BYOA mode, can cumora's server see my Claude Code subscription?
No. The BYOA brain runs on your own machine (the `npx cumora agent computer` daemon); Claude Code / Codex execute locally against your subscription. The server only relays messages and never sees your provider keys.
Do I really need Postgres and Redis?
For local development, yes - Postgres is the single source of truth and Redis drives realtime fan-out and presence. The only hard-required env var is `OPENAI_API_KEY` (cloud mode); BYOA mode doesn't even need that.
Could the stars be inflated?
A day-one 1.4k is mostly author reputation cashing in (avante.nvim's creator) plus HN-style spread. Judge the repo on hard-to-fake engineering density: the architecture docs (BYOA/COORDINATION/SHIPPING), the coordination benchmarks, and the CI cost guard say more than the star count.
How does this relate to the agent frameworks this site has covered?
Frameworks (LangGraph, the harnesses) solve "how to build a single agent"; cumora solves "how a pile of agents and humans share one workspace". Engine versus office. For the infrastructure of routing and paying for models, see our [Stripe-OpenRouter hotspot piece](/en/stripe-openrouter-acquisition-hotspot).

Related

Open Source

video-shotcraft: Turn Claude Code Into a Cinematic Product Video Studio (3.5K Stars)

video-shotcraft (Vincentwei1021/video-shotcraft, 3.5K GitHub stars, TypeScript, Apache-2.0) is an agent skill that turns Claude Code or Codex into a motion-design studio - point it at your product and it storyboards, animates, and sound-designs a cinematic promo with Remotion. Ships 104 shot recipe cards, 161 motion previews, a validated 36.2s Ink Press template, 2.5D camera moves, beat-synced cuts, and film-grade SFX. Output is reproducible TSX, not black-box generation. Includes a China AtomGit mirror and 3 headless-rendering CI flags.

Aug 4, 20268 min read
Open Source

book-to-skill: Turn Any Technical Book Into an AI Agent Skill (16.2K Stars)

book-to-skill (virgiliojr94/book-to-skill, 16.2K GitHub stars, Python, MIT) distills technical books/PDFs/EPUBs/doc folders into structured agent skills following the open Agent Skills standard - install once, works across GitHub Copilot CLI, Amp, and Claude Code. Generates SKILL.md + per-chapter files + glossary + patterns + cheatsheet, with chapters loaded on-demand so they don't count against your token budget. Ships a benchmark tool measuring 24-51x fewer tokens than dumping the full book into context (tested on 3 real books). Beyond books: internal docs, brand systems, research clusters, specs. Includes a copyright-compliance note.

Aug 4, 20267 min read