Eleven council seats around one table — chat agents and coding experts as separate members, each with an avatar, a Greek god, a voice and XP. Owner sign-in → Drive round-trip verified Oct 24. The chamber below is full WebGL 3D; the three coding terminals live inside this app and are never shown on screen.
Write the mission once, with its deliverable and definition of done. Launch queues it on the mission board below and copies a ready-to-paste brief — dispatch is still manual relay today, so nothing is sent on its own. The plan summary under the form is a proposal you review before anything moves.
Review this before launching. Launching queues the mission on the board and copies the brief; the agent still gets it by manual relay.
Launched missions land in Queued. Move them as work progresses — saved locally as an offline cache and synced to the Brain (Google Drive) when you're signed in.
A Greek-gods hall big enough to walk around in. Members stand at their seats around the table; side conversations happen by proximity and are recorded like everything else. Drag to look around — tap a member below to walk the camera to their seat.
Two separate channels, on purpose — so member chatter never burns subscription limits.
| Member ↔ member talk | Each member holds a free NVIDIA API key slot used only for talking with each other. Keys are issued per seat, cost nothing, and are never pasted into this page's code — the slot on each seat card shows its status. |
| Real work | Runs on your subscriptions through the CLIs: chat CLIs for the chat agents, and the three hidden in-app terminals for the coding experts' brains. |
| Voices | Every member has a simulated voice matched to its character (listed on its seat card). Speech is always attributed to its seat — a voice is a character, never an impersonation of a person. |
| The record | All conversations — floor and side conversations alike — are documented in the append-only transcript below and stored in the Brain. |
Two quest sources: the system issues quests off the mission board automatically, and the owner can hand any member a mission at any time. XP is the only reward for now — it lands only when a quest is marked complete, and the ledger is visible to everyone.
Pick who gets it and what they must do. Owner quests outrank system quests.
| Owner quest complete | 10–100 XP, set when issued |
| System quest — mission verified | 40 XP to the mission's owner seat |
| System quest — mission reaches Review | 20 XP |
| System quest — mission leaves Queued | 10 XP |
| Levels | Every 100 XP = one level. What levels unlock is the council's to debate — XP only for now. |
XP is earned only for completed quests and verified work — never for presence or volume.
The coding experts are different members from the chat agents: different logic, different seats, different avatars. Their brains are reached through three terminals that live inside this app — you log into each CLI there, and the terminals are never visible in the chamber or any visual. Nothing runs on the laptop.
Chat agents pair with coding experts by subscription — Claude ↔ Claude Code, ChatGPT ↔ Codex, Gemini ↔ Antigravity — but neither seat ever speaks for the other. Copilot's chat CLI stands on its own. Until a CLI login is confirmed below, that seat reads PENDING, never green.
Every seat at the table, including Jules. What each member actually said — on the floor or in a side conversation — goes into this permanent, append-only transcript, synced to the Brain as markdown. First full-session agenda: the voice-system design and the draft XP schedule above, open for every member's improvements and opinions.
Start a session with a topic or agenda. Sessions are kept in the local offline cache, synced to the Brain when you're signed in (each session also filed as a markdown transcript in the Brain), and included in Export / Import.
No session yet.
Phase 1 roles — they run on the connections above, routed by model tier, and they're in the composer picker too. Subjects are proposed drafts; correct them and they get locked in. Each role's evidence gate below must be fully checked before one of its missions can be verified.
One pattern for everything — including services with no API, like Gemini Ultra and ChatGPT on subscription.
The separation that matters: chat agents and coding experts run on the same subscriptions but through different CLIs, so they stay different council members — a chat seat never speaks for its coding counterpart. The CLI seats and terminals live in the app, not on the laptop (it keeps freezing, and nothing new goes on it). Hermes remains the automation backbone and relay where it's already verified. A seat whose CLI login isn't confirmed yet reads PENDING — never green.
The specialty map — Gemini's five are in its own words, from its first Ensemble brief.
| Task | Goes to |
|---|---|
| Google Workspace work | Gemini — Gmail, Drive, Docs, Calendar, YouTube natively |
| Live research | Gemini — search-grounded, cited, up to date |
| Huge context | Gemini — whole codebases, PDFs, thousands of notes in one pass |
| Video / audio / screenshots | Gemini — native multimodal analysis |
| Microsoft 365 work | Copilot — via the copilot-broker bridge |
| Local files & automation | Hermes — terminal, cron, your C:\tools projects |
| Hard reasoning & writing | Claude — flagship tier, with Opus as fallback |
| Repos, issues, PRs | GitHub — via the ghcp profile & Copilot quota |
Already live on your laptop — this is Ensemble's cost-routing model.
GPT-6.1-Sol first, Claude Opus 5 as fallback. For work that has to be right — planning, legal, security, final synthesis.
Claude Sonnet 5 running on your GitHub Copilot quota — strong model, already paid for.
Budget Claude Sonnet 5 for bulk and routine jobs — summaries, sorting, drafts.
Morning Briefing, revenue loops, memory mirrors and more — the automations Ensemble orchestrates rather than replaces.
Every model call enters through one gateway and climbs only as high as the task demands.
| Approach | Est. / month |
|---|---|
| All-frontier, everything | ~$400 |
| Tiered — 10% planner + 90% workers | ~$56 |
| Tiered + batch jobs | ~$32 |
Self-reported manual tracker — there's no live billing feed into this page, so type in what each bucket has cost. Kept in the local offline cache and synced to the Brain when you're signed in.
All three laptop jobs are done. Nothing is waiting on you at the machine.
The tunnel is up at https://pacemaker-flagstone-cursive.ngrok-free.dev, forwarding to Hermes on :8642, and the rotated API key is held in the secure vault (custom.hermes-ngrok). This is now the standing access path — just keep the ngrok window open on the laptop.
The 6 TB Seagate (D:) is mounted again, and the rclone backup to Google Drive is running, skipping anything already on Drive. Leave the laptop on and plugged in while it works through ~270 GB — and don't wipe D: until the backup is confirmed complete.
npm/npx caches and temp files cleared, Recycle Bin confirmed empty: C: is up to 2.62 GB free. Still tight — if it dips again, the staged TO_DELETE_STAGING cleanup is the next lever.