Do I need to keep the app installed?

Install the lightweight racn-node peer once. After that, chat at metademic.org/chat from any browser — jobs route to the P2P mesh (yours or others). No need to keep a window open.

500 credits on signup. Each prompt spends ~10 credits. Contribute GPU work and earn earn = 10 × (1 + 0.1·trust) on verification. Confidential tasks use trust ≥0.9 and NaCl SealedBox.

Choose per prompt: Public / Protected (trust≥0.7) / Confidential (trust≥0.9 + sealed) / Local-only (never leaves your machine). Coordinator never sees plaintext of confidential jobs.

RACoN · Reciprocal Agentic Compute Network

Turn idle GPUsinto shared supercompute.

RACoN is hybrid: install the peer node once on your machine — then do everything from the web. Chat at /chat, history + credits in Supabase, compute on the P2P mesh: pipeline (sequential) or mesh (DAG: retrieval/code/doc → verification → synthesis), capability-aware scoring, 18 s lease + 5 s monitor.

Download node Open web chat How it works

Real system — not a mock

From your Python prototype to the web

Control-plane (FastAPI :8443) + coordinator + Redis queue/lease/inflight + MinIO shards. Workers register → heartbeat 15 s → POST /prompt → planner → scheduler → WSS job frame → worker decrypts → generate(prompt, max_tokens) → signed complete → ledger earn.

Why hybrid?

Web = identity, chat history, credits, billing (Supabase Auth + RLS).
Peer = your GPUs, your data, local-only option never leaves.
Install once → use from any browser, any machine.

Download Try chat

Credits

Bootstrap 500 on signup. Cost ~10 per prompt. Earn 10 × (1 + 0.1·trust) on verification (2-of-3 sampling for high value). 402 if insufficient. Tables: credits, credit_ledger (RLS per user).

How it works

Planner · Scheduler · Lease

Prompt → pipeline (≤20 words: model_forward → 2-stage high→medium) or mesh (document-analysis / code / retrieval → verification → synthesis). Scheduler scores w_g·gpu + w_m·mem + w_b·bw − w_l·lat − w_q·queue + w_shard·locality + w_trust·trust, privacy gates (confidential ≥0.9, protected >0.7). Lease 18 s, monitor 5 s, requeue on miss.

Torrent shards

cfg → grant → data → finish. Adaptive engine: vLLM / llama.cpp / MLX.

Confidential

Ed25519 + X25519, NaCl SealedBox to worker box_pubkey, signed POST /worker/complete.

Self-healing

Heartbeat extends lease; monitor requeues if worker !active and now − assigned_at > lease. MAX_ATTEMPTS 3.

Field notes: distributed inference →·Install guide →

Quick start

Copy, run, chat

pip install racn-node
racn-node init --coordinator-url wss://coordinator.metademic.org/ws
racn-node start
# then open https://metademic.org/chat

Outbound WSS only — NAT-friendly, TURN relay via ws_client.py. H1 (16GB GPU) → H6 (CPU/Apple Silicon).

APIs the site uses

POST /prompt → job_id · privacy_tier · node_id
GET /job/:id → status/output · polled from /api/racn/job
WSS /ws · auth→hello(JWT)· register · heartbeat · job · complete · relay
Supabase: profiles, credits, racn_nodes, conversations, messages, racn_jobs

Start chatting Download node Open Research →