Real-time brainstorming between several AI models. Pose a topic with optional evidence; the app drives multiple OpenRouter models through a three-phase discussion — parallel race → closing reflections → consensus wiki — and streams the whole thing to a single-page frontend.
Each model runs its own independent loop:
- It fires a chat-completion request to OpenRouter on a snapshot of the transcript at that moment.
- Whichever model finishes first appends its reply to the shared transcript.
- Slow replies are not cancelled — they land later, on top of whatever the
other models said in the meantime. The
stale_byfield on each message records how many messages had appeared between the snapshot and the append. - After landing a message, a model must wait
ceil(N/2)other messages before it can fire again (the cooldown), whereNis the number of participants. - Replies are hard-truncated to
char_limit + 100characters; the system prompt asks forchar_limitto keep them concise. - Models are encouraged to address each other with
@Nametags, which the UI highlights in the addressee's color.
Once the race ends, every model that spoke is asked, in parallel:
Have you revised your initial position?
Each reflection starts with a tag of the form [shift: N] (0–100, where 0 =
unchanged, 100 = completely revised). The server parses the tag and the UI
renders a small bar in the model's color showing the magnitude of the shift.
The models then crowdsource a shared summary, wiki-style:
- A random first editor proposes the initial draft.
- For each remaining model in random order, that model proposes ONE focused
revision (an addition, deletion, rephrasing, or correction). The full
revised article comes back via a
SUMMARY: …\n===ARTICLE===\n<body>delimiter protocol; JSON and bare-prose fallbacks keep parsing robust. - The other models vote in parallel:
yes,no, orabstain(with a one-sentence reason). Votes stream into the UI one at a time. - The revision is applied iff
1 + yes_count ≥ N / 2— the proposer's implicit yes plus enough explicit yeses from the rest. - The current draft is updated and the next proposer takes a turn.
The UI shows a live diff (line-level, +/− highlighted) for every proposal so you can see exactly what was added or removed, plus a final "Consensus summary" panel and a tally of accepted vs. rejected revisions.
The initial draft is capped at char_limit × 3 characters; revisions from
turn 2 onwards have no character cap.
Every discussion is auto-saved as it runs:
discussions/{timestamp}-{slug}-{shortid}.md— full transcript (messages + reflections with shifts + consensus turn-by-turn log with diffs and votes).discussions/{base}-consensus.md— companion file with just the final consensus summary, written once at the end of phase 3.
Files are refreshed after every message and turn, so a crash mid-run still
leaves a partial transcript on disk. The discussions/ directory is
gitignored.
python3 -m venv .venv
.venv/bin/pip install -r requirements.txt
cp .env.example .env # then add your OPENROUTER_API_KEYGet an OpenRouter key at https://openrouter.ai/keys.
.venv/bin/uvicorn server:app --reloadOpen http://localhost:8000.
OPENROUTER_API_KEY(required) — set in.envor the environment.OPENROUTER_REFERRER(optional) — sent as theHTTP-Refererheader to OpenRouter.- Default model list lives in
DEFAULT_MODELSinserver.py. Override per request from the UI. - Model icons live in
frontend/icons/and are mapped to model ids inMODEL_ICONS(and colors inMODEL_COLORS) insidefrontend/index.html.
| Field | Meaning | Default |
|---|---|---|
topic |
The question or prompt to discuss | required |
context |
Free-form evidence / background passed to every model | "" |
char_limit |
Hard cap on the length of each phase-1 reply (chars) | 600 |
max_turns |
Total fires across all participants in phase 1 | 12 |
models |
List of {id, name} from OpenRouter |
all defaults |
Cooldown is derived: min(ceil(N/2), N - 1). The wiki article cap on the
initial draft is char_limit × 3.
GET /api/health—{ ok, openrouter_configured }GET /api/models— default model listPOST /api/discussions— start a discussion; returns{ id, models, ... }GET /api/discussions/{id}— current state and historyGET /api/discussions/{id}/stream— SSE stream. Event types:- Phase 1:
start,speaker_started,message,speaker_failed - Phase 2:
closing_started,reflection,reflection_failed - Phase 3:
consensus_started,consensus_proposing,consensus_voting,consensus_vote,consensus_turn,consensus_completed,consensus_failed - Terminal:
completed,error
- Phase 1:
- Discussions live in memory only; they vanish on restart (transcripts remain on disk).
- The initial burst means all
Nmodels reply once on an empty transcript, so the firstNmessages are independent opening statements. - Every fire bills against OpenRouter — slow models that arrive late still cost what they cost.
- Phase 3 makes one proposal call per model plus
(N − 1)vote calls per proposal turn, so total cost roughly scalesO(N²). - If you want stale phase-1 replies dropped instead of appended, add a check
in
speaker_loopagainstlen(disc.history) - snapshot_lenbefore appending.