{"id":51462040,"url":"https://github.com/langwatch/tokenmaxxer","last_synced_at":"2026-07-06T06:31:27.156Z","repository":{"id":364580923,"uuid":"1267703494","full_name":"langwatch/tokenmaxxer","owner":"langwatch","description":null,"archived":false,"fork":false,"pushed_at":"2026-06-13T17:03:11.000Z","size":483,"stargazers_count":4,"open_issues_count":0,"forks_count":0,"subscribers_count":0,"default_branch":"main","last_synced_at":"2026-06-13T17:07:30.957Z","etag":null,"topics":[],"latest_commit_sha":null,"homepage":null,"language":"TypeScript","has_issues":true,"has_wiki":null,"has_pages":null,"mirror_url":null,"source_name":null,"license":null,"status":null,"scm":"git","pull_requests_enabled":true,"icon_url":"https://github.com/langwatch.png","metadata":{"files":{"readme":"README.md","changelog":null,"contributing":null,"funding":null,"license":null,"code_of_conduct":null,"threat_model":null,"audit":null,"citation":null,"codeowners":null,"security":null,"support":null,"governance":null,"roadmap":null,"authors":null,"dei":null,"publiccode":null,"codemeta":null,"zenodo":null,"notice":null,"maintainers":null,"copyright":null,"agents":null,"dco":null,"cla":null}},"created_at":"2026-06-12T19:36:20.000Z","updated_at":"2026-06-13T16:19:17.000Z","dependencies_parsed_at":null,"dependency_job_id":null,"html_url":"https://github.com/langwatch/tokenmaxxer","commit_stats":null,"previous_names":["langwatch/tokenmaxxer"],"tags_count":null,"template":false,"template_full_name":null,"purl":"pkg:github/langwatch/tokenmaxxer","repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/langwatch%2Ftokenmaxxer","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/langwatch%2Ftokenmaxxer/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/langwatch%2Ftokenmaxxer/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/langwatch%2Ftokenmaxxer/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/langwatch","download_url":"https://codeload.github.com/langwatch/tokenmaxxer/tar.gz/refs/heads/main","sbom_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/langwatch%2Ftokenmaxxer/sbom","scorecard":null,"host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":286080680,"owners_count":35180933,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2026-05-26T15:22:16.424Z","status":"online","status_checked_at":"2026-07-06T02:00:07.184Z","response_time":106,"last_error":null,"robots_txt_status":"success","robots_txt_updated_at":"2025-07-24T06:49:26.215Z","robots_txt_url":"https://github.com/robots.txt","online":true,"can_crawl_api":true,"host_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub","repositories_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories","repository_names_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repository_names","owners_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners"}},"keywords":[],"created_at":"2026-07-06T06:31:26.218Z","updated_at":"2026-07-06T06:31:27.139Z","avatar_url":"https://github.com/langwatch.png","language":"TypeScript","funding_links":[],"categories":[],"sub_categories":[],"readme":"# tokenmaxxer\n\n**The meeting room that ships while you talk.**\n\nHackathon theme: *what if compute and tokens were no longer a limitation?*\n\nYou sit down for a day of work. Microphones are on. Max, a realtime voice\nagent, is in the room with you. You say \"get five agents on a full dark mode\nfor the website\" and five Claude Code agents spin up, pile into one shared\nchat channel, and self-organize, one claims the CSS foundations, another\ntakes the pages, a third reviews. You didn't assign roles. You keep talking;\nthey keep working. Compute is unlimited, hesitation is the only cost.\n\n**And it answers in about half a second.** You finish a sentence and the\nterminals are already opening. The realtime voice loop (Inworld gemma-4) runs\nat roughly **500ms** from your voice to the swarm moving, so it feels like\ntalking to someone in the room, not typing a command and waiting on a\nspinner.\n\n![Three agents splitting a dark-mode mission between themselves in one shared channel, from a single spoken sentence, with no roles assigned by a human](https://raw.githubusercontent.com/langwatch/pr-screenshots/main/pr-1/swarm-self-organizing.png)\n\n*One spoken \"get a few agents on dark mode\" and the room fills with named\nagents that divide the work themselves in the channel: one takes the CSS\nfoundations and theme toggle, another claims the pages, down to the ~89\nhardcoded colors. Nobody assigned a thing.*\n\n## It controls your machine, not a tab\n\ntokenmaxxer is not a webpage you build inside, it drives your real desktop.\nSay \"spin up a room to fix the login\" and real terminals fly open with Claude\nCode agents working in them. The KanbanCode board comes forward showing the\nagents talk to each other in real time. Say \"pull up the website\" and a real\nbrowser window opens at the running site. A floating HUD shows the room is\nlistening, on top of everything.\n\n```\n room mic ──► HUD overlay (Electron) ──ws──► gateway (server, on your Mac)\n              always-on-top, transparent          │  ←ws→ Inworld Realtime (gemma-4)\n                                                   │\n   ┌── desktop control (osascript / Warp) ◄────────┤ tool calls\n   │     • spawn_room → N Warp terminals, tmux, Claude Code, watch them work\n   │     • focusChannel → KanbanCode board forward on the room's channel\n   │     • open_url → chromeless Chrome window at the live site / a GitHub issue\n   │     • done → macOS notification + the PR window, proactively\n   │\n   └── the swarm appears AROUND the HUD as real windows\n                                                   │\n   room ◄── kanban channel (the agent loop) ── kanban CLI ── tmux ── N × claude code\n            agents claim slices, review each other, coordinate\n```\n\n### Why this wins (the swarm, not a single hidden agent)\n\n- **Realtime, not request-and-wait**: ~500ms from your voice to the swarm\n  moving. You steer N agents by talking over them, live, the way you'd brief a\n  team, never a type-a-prompt-and-wait loop.\n- **Agentic leverage**: N independent Claude agents on one mission, plus a\n  voice agent listening live to steer them. Not one agent with hidden\n  sub-tasks: N real sessions you can see and talk to.\n- **Measurable impact**: 1 agent is one context window on one thread. N\n  agents is N context windows working the problem in parallel. Separate\n  sessions give you full control of all of them.\n- **The agent loop**: the shared channel is a natural ralph-loop: agents\n  keep messaging each other, finding issues, and pushing forward. One can\n  step up as supervisor. No better loop than agents holding each other to it.\n- **Cost / quality**: because the agents are under your control and visible,\n  they don't burn tokens mysteriously doing random things.\n- **Collaboration**: the team guides the swarm by voice, live, together.\n- **Technical ambition**: the desktop layer is built behind a\n  platform-agnostic adapter so it runs on Windows too, not just macOS.\n\n## The pieces\n\n- **server/**: voice gateway (OpenAI Realtime-compatible WS, proxies\n  Inworld), server-side tool execution, the **room engine**\n  (`orchestrator/room.ts`: spawn a swarm into a kanban channel, the\n  coordinate-first agent brief, a deterministic project registry), and the\n  **desktop control layer** (platform-agnostic `DesktopController`; macOS\n  adapter via osascript + Warp + chromeless Chrome; Windows stub for the\n  incoming teammate).\n- **hud/**: Electron shell, a frameless, transparent, always-on-top overlay\n  wrapping the console in HUD mode (reuses the console's mic/audio/WS stack).\n- **console/**: the room console (transcript, tool feed, live fleet) and the\n  compact HUD view (`?hud=1`).\n- **specs/**: BDD feature specs for everything above.\n\nThe agents talk over [Kanban Code](https://github.com/langwatch) channels:\neach agent has a tmux session kanban tracks, and a `kanban channel send`\nbroadcasts into every other agent's pane. Max joins the channel too (as\n`@max`) so he can drop the mission in and pass along corrections by voice.\n\n## Run it\n\n```bash\npnpm install\ncp .env.example .env   # INWORLD_API_KEY, ANTHROPIC_API_KEY, OPENAI_API_KEY,\n                       # LANGWATCH_API_KEY\npnpm dev               # gateway :4870 + console :5170\npnpm hud               # the floating always-on-top listening HUD (Electron)\n```\n\nDesktop control is on by default on macOS (`TOKENMAXXER_DESKTOP=0` to run\nheadless). Click **listen** on the HUD and talk. Demo script and failure\nplaybook: [DEMO.md](DEMO.md).\n\nThe agents spawn through the kanban CLI in tmux; `TOKENMAXXER_AGENT_MODEL`\npicks their model (`sonnet` for testing). The project a room works in is\nresolved from what you say (\"the website\" → the local site checkout,\n\"langwatch\" → the langwatch repo) in `server/src/orchestrator/projects.ts`.\n\n## Test it\n\n```bash\ncd server\nnpx vitest run tests/room.unit.test.ts tests/desktop.test.ts   # network-free, CI\npnpm typecheck\n```\n\nThe network-free unit tests cover the engine's pure core: how a spoken hint\nresolves to a project, how a topic becomes a channel name, and the\ncoordinate-first brief every agent receives (it must tell the agent to JOIN\nthe channel and coordinate BEFORE building, that ordering is the whole\ndelegation contract). The scenario voice tests and the delegation experiment\neval drive the real gateway with audio and a judge; they need live APIs, the\nkanban CLI, and tmux, so they run locally, not in CI.\n","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Flangwatch%2Ftokenmaxxer","html_url":"https://awesome.ecosyste.ms/projects/github.com%2Flangwatch%2Ftokenmaxxer","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Flangwatch%2Ftokenmaxxer/lists"}