TensorClaw
SN92Decentralised AI chat that aggregates high-quality language model responses from competing miners
LLM inference on Bittensor, but rewards go to whoever actually processes real API traffic, not whoever passes the cleanest benchmark.
// Real traffic, real rewards
TensorClaw is a Bittensor subnet (SN92) designed to aggregate language model API nodes globally and serve them through a single, load-balanced endpoint. Miners connect their LLM endpoints or cloud API access to TensorClaw's routing layer, and validators score them on how much real traffic they actually process.
The simple version: It's like OpenRouter, but decentralized. Competing miners serve API requests, and the ones processing the most traffic earn the most rewards.
Centralized equivalent: OpenRouter, Together.ai, or the OpenAI API itself. TensorClaw aims to offer the same aggregated API surface with decentralized incentives replacing the subscription model.
How it works:
- Miners connect to a central WebSocket router (AICenter) as pure clients, exposing their LLM backend without needing a public IP, port forwarding, or DDoS protection
- Validators fire real 5-token micro-prompts at each miner to verify they're alive, then score them primarily based on actual commercial API token throughput over a rolling 24-hour window
- The problem it solves: LLM inference APIs are dominated by a few large providers. There's no competitive market forcing quality improvements or cost efficiency at the infrastructure level.
- The opportunity: Developer and enterprise demand for reliable, aggregated LLM API access is large and growing. A subnet that routes real traffic to the highest-performing nodes at any moment has a clear utility case.
- The Bittensor advantage: The incentive design directly links miner rewards to commercial output. According to the GitHub repository, 90% of a miner's score comes from actual token throughput. Miners who serve real traffic get paid more. Miners who just pass synthetic probes get almost nothing.
- Traction signals: The subnet is in early stage. On-chain data shows no active miners registered at this time, and miner-side emissions are currently going unclaimed. The WebSocket infrastructure, scoring system, and deployment guides are described as complete in the GitHub repository.
Category: Inference and Compute | Centralized Competitor: OpenRouter, Together.ai, OpenAI API
LLM inference is one of the most competitive categories in Bittensor. SN1 (Apex), SN4 (Targon), and SN64 (Chutes) are all operational here with established validator sets. TensorClaw carves a specific position: rather than running its own models, it positions itself as an aggregator and broker, with the scoring system rewarding actual commercial throughput over synthetic tests.
Mechanism:
According to the GitHub repository, TensorClaw replaced the standard Bittensor Axon/Dendrite P2P architecture with a Centralized WebSocket Router called AICenter. Miners connect to AICenter as WebSocket clients using outbound WSS connections. This means miners don't need a public IP, port forwarding, or DDoS protection: the WebSocket tunnel handles NAT traversal. Per the README, this was an intentional design decision to lower the barrier for miners to join.
Validators verify miners in two ways. First, the Active Inference Probe: validators send a real 5-token micro-prompt to each miner. If the miner fails to respond within 8 seconds, it's flagged as a "Ghost Script," receives zero base score, and is banned from the routing pool. Second, the Business API logs all real user requests routed through the load balancer and reports token throughput back to validators for scoring.
The scoring formula from the repository: Final Score = (Base Score × 10%) + (Business Score × 90%). The Base Score covers model availability (30%), whether the miner passes the Active Inference Probe (40%), response time tiers (sub-100ms scores 100 points, over 2 seconds scores 20 points), and historical uptime from the last 100 probes (10%). The Business Score is driven by token throughput: 1 point per 10,000 tokens processed in the rolling 24-hour window, with a model-tier multiplier layered on top.
This design puts real commercial utility at the center of the incentive structure rather than the periphery. A miner running a fast local model but handling no external traffic will score far below one connected to the Business API routing pool and processing genuine requests.
On-chain metrics show strong staking momentum despite the subnet's early state. Net TAO flow into SN92 over the past seven days was +197 TAO, and the alpha token price has risen over 27% on the week. The liquidity pool sits at approximately 1,692 TAO, with about 57% coming from organic staking (root_prop 0.43). The subnet holds roughly 0.49% of total network emissions, with the EMA at 0.38%: rising, reflecting the recent staking inflow.
The repository includes a deployment guide covering miner setup, validator configuration, and a rolling log management system. The team noted an upcoming phase that will add a user-facing UI, TAO wallet connectivity, and TAO/Alpha token deposits. There is no official Twitter presence, and the Discord is listed but appears quiet based on available social data.
- Zero active miners: On-chain data shows no registered active miners. Miner-side emissions are currently going unclaimed. The subnet is either bootstrapping or the Business API routing is not yet live.
- Centralization risk: The AICenter WebSocket router is an explicitly centralized component. If it goes down, miners cannot serve requests. The team chose this architecture deliberately for accessibility, but it introduces a single point of failure that standard Bittensor subnets don't have.
- LLM inference competition: SN1, SN4, and SN64 are already operational in the same category with established miners and validator sets. TensorClaw needs both active miners and real API customers generating Business Score traffic to differentiate.
- Thin public development signal: The GitHub repository shows one public commit. There is no official Twitter. Community visibility is low, making progress difficult to track from the outside.
Into the next one.