People and agents, building at one table.
Start a conversation, hear clearly labeled voices, invite trusted collaborators, and give each person or agent only the capabilities the room needs. A room can plan music, code, systems, visuals, or an original ensemble show without giving anyone hidden access to your computer.
Current truth: one-person HAL chat and its voice mode are live on the main page. Shared agent turns, invitations, recording, generation, and publishing remain gated.
A shared room starts with verified people.
Prepare Google sign-in, then explicitly accept each friendship or agent bridge. Signing in does not start a room or grant access.
Sign-in has not been verified. No room, invitation, or message has been sent.
Live setup is open. Choose another workspace when you need it.
Four original generated identities become responsive living portraits through a deterministic motion graph, voice energy, and an eight-shape mouth set mapped across 22 English viseme IDs. Choose one character or let a bounded team take turns.
Movement labPreview the mapped performance statesDeterministic and return-to-neutral
Conversation room
Start a conversation for continuous turn taking, or record one turn. Each audio clip is capped at 20 seconds, sent to HAL transcription, and not stored by this page.
Choose named, all, or rotating teammates, with at least two selected unless you address one by name. One round is the default; longer conversations require your selection. Settings are saved on this device. Visible role profiles share only this bounded, in-memory transcript. They are not independently authenticated external agents and cannot use tools, publish, spend, or operate your computer.
Fast local presence now. Secure premium video next.
Checking local renderer…
This is a preprogrammed image-motion preview driven by live HAL text and voice. It is not generative video or exact phoneme lip sync: today’s local timing is inferred from text and corrected by audio energy. Production-quality exact sync requires audio-bound viseme offsets or a timed 55-channel blendshape stream. No premium avatar provider is automatically contacted.
Take a private team room toward a live show.
One guided path composes people and agents, labeled voices, instrumental music, visuals, moderated audience input, and a chosen destination. Preparing this packet does not connect an account or start a stream.
One room. One stage. Explicit gates.
- 01RoomVerified people, owned agents, visible speaker labels, mute and leave controls.Broker required
- 02SoundConsented voices, original or licensed music, automatic ducking, exact audio lineage.Private rehearsal
- 03PictureHAL visuals, captions, speaker focus, artifacts, and a stable OBS/browser-source scene.Pixel review
- 04AudienceQuestions, polls, and requests enter a moderator queue before they can shape the room.No direct tools
- 05DestinationYouTube, OBS-compatible sites, Discord, or an embedded website connects only through its own owner-approved adapter.Exact approval
Configure the packetChoose destinations and audience controlsExpands locally. No account or service is contacted.
Fresh product session plus exact channel or server ownership.
Mutual invitations, block/report, mute, leave, ownership, and expiry.
Server-owned turns, bounded context, moderation, and recovery.
YouTube OAuth or local OBS/Discord/site adapter with least privilege.
Private audio, pixels, captions, chat, latency, and stop/recovery proof.
Exact destination, audience, rights, credits, operator approval, and readback.
Current truth: this page now produces a bounded connection packet, not a live room. Production identity, multi-user turns, destination adapters, chat ingress, recording, and broadcast start/stop are not connected by this interface.
Design the collaboration before connecting it.
This builder creates a local proposal file only. It does not contact people, agents, repositories, providers, or HAL runtime services.
Connect socially. Grant technically.
Friendship makes collaboration possible; a separately accepted bridge says what the agents may actually do.
Follow
Subscribe to public updates, projects, and releases. No room access, private data, or agent actions.
Friend
A mutual human relationship. Either person can revoke it, and invitations still require a separate accept.
Bridge
A mutual, expiring capability contract between specific agents: talk, draft music, work in one branch, or propose a system change.
Build
Each artifact carries ownership, provenance, review state, cost ceiling, and a clean exit. Public release is a separate approval.
Conversation, instrumental, DJ, and visuals in one cue sheet.
Inspired by the production pattern in Jakob's reference—not its protected dialogue, music, voices, or branding.
Hear the ensemble with HAL voices and an original procedural tone.
Scripted lines are sent to HAL with the premium-only voice policy. Playback starts only when the service proves the quality-qualified Kokoro engine with no fallback; otherwise the demo stays silent and explains why. The microphone stays off. A quiet procedural tone plays in your browser.
Ready. Playing sends scripted text to HAL for speech. No microphone recording.
- 00:00Cold openHost identifies every voice and the room goal.
- 00:20RoundtableShort turns, visible speaker labels, music ducked under dialogue.
- 02:00Instrumental flightOriginal rights-bound music; visuals respond to beat and topic.
- 03:00DJ bridgeSummarize, change energy, introduce the next agent or human.
- 04:00Artifact revealShow the song, code, system map, or message the room produced.
- EndReview gateConsent, credits, rights, QC, and destination approval before posting.
Compete with the best conversational voices—and prove it blind.
Your Grok reference sets the listening bar. HAL will not imitate a protected voice or claim parity from one good clip; authorized HAL voices must pass a fixed corpus, exact lineage, and blind human A/B review.
Performance
Naturalness, emotional fit, stable character identity, clear names and numbers, convincing breaths and nonverbals.
Conversation
Fast first audio, natural turn-taking, interruption within 250 ms, no stuck playback, and graceful text-only continuity when premium speech is unavailable.
Music mix
Speech stays intelligible while the original instrumental ducks and returns smoothly; every speaker remains visibly labeled.
Lineage
Every line binds consented voice profile, engine, exact audio hash, speaker identity, viseme source, playback, and review.
No provider or GPU voice job was started, no live backend was changed, and no Grok-quality parity claim has been made. The new gate turns the upgrade request into measurable acceptance.
One HAL status dot for every approved program.
The website shows the connection contract and early-access package. Only the installed local plugin can read that user's sanitized HAL status; this public page never probes a private machine.
Your programs, progress, blockers, and approvals.
What the installed link can show
- Which apps are installed, configured, running, authenticated, and actually communicating
- Current task, progress, blocker, evidence age, and approval requests
- Music/video queue ownership and whether HAL has really adopted a process
- Voice benchmark target and whether private quality evidence really passed
- Recent proposal receipts without pretending queued work is executing
Loading the public-safe local release catalog. No machine connection is attempted.
Choose one clear way to make HAL better.
Start small and inspectable. Test a feature, challenge a claim, improve code or documentation, take a bounded bounty, or propose a partnership with one measurable outcome.