SOMA
SN114AI tools delivered through MCP infrastructure, plug-and-play solutions for developers
A weekly tournament for MCP servers. Miners ship algorithms, validators score them against a moving target, and one winner takes the cycle.
// MCP infrastructure, weekly winners only
SOMA brings Model Context Protocol (MCP) servers into Bittensor. MCP is the standard that lets AI agents call external tools, query data sources, and run code safely. SOMA turns that infrastructure into a competitive market, where the best implementation each week earns the subnet incentive.
The simple version: It's like a weekly hackathon for AI plumbing, except the leaderboard pays the winner in TAO.
Centralized equivalent: Hosted MCP catalogs and curated tool registries inside the broader Anthropic MCP ecosystem, LangChain, and managed AI gateways like Cloudflare AI.
How it works:
- Miners submit one algorithm per hotkey per week. The platform screens for stability, then qualifies the top performers for the live competition phase.
- Validators fetch execution results from the SOMA platform, score them against the active task criteria, and report scores back. The platform aggregates the scores and validators set on-chain weights from the platform's rankings.
- The problem it solves: MCP servers exist, but there is no trustless way to rank them by quality or pay for the best implementation. Most live behind hosted services with no public benchmark and no shared incentive layer.
- The opportunity: If agents are going to depend on tool calls for context, memory, and execution, the underlying MCP services need a market that pays for uptime, latency, and quality, not just for existing.
- The Bittensor advantage: A weekly winner-takes-all cycle forces continuous iteration. Stagnant solutions get replaced; new algorithms can win their first week. The same incentive treadmill that produced competitive inference subnets is now pointed at agent infrastructure.
- Traction signals: The team shipped SOMARIZER, a beta context compression tool that uses the subnet, in April 2026. The second competition, Agent Chain-of-Thought Compression, launched May 4. An integration with @openclaw is in progress, and the team posted a public roadmap pointing to a first marketplace MCP going live in Q3 2026.
Category: Inference and Compute | Centralized Competitor: Anthropic MCP ecosystem, LangChain, Cloudflare AI
MCP is Anthropic's standard for letting models talk to external systems. Adoption has moved fast on the client side, with most major agent frameworks now speaking MCP, but the server side is fragmented. Quality varies, hosting is centralized, and there is no objective benchmark for "this MCP server is better than that one." SOMA is a bet that the server side becomes the bottleneck once agents are running real workloads, and that a Bittensor-style market is the right way to sort it out.
Mechanism:
The subnet runs on a weekly competition cycle with three phases. During the submission window, miners upload algorithms tied to a registered hotkey on netuid 114. A screening phase runs automated validation and integrity checks, then promotes the top qualified submissions. In the competition phase, validators evaluate those submissions continuously against the active task criteria and report scores to the SOMA platform. At cycle close, the platform aggregates rankings, the top-ranked miner takes the subnet incentive allocation, and a new week begins.
The first task is context compression: shrinking prompts and conversation history without degrading model output. Lower token counts mean lower inference cost and longer effective memory for agents. The second task, currently live, is Agent Chain-of-Thought Compression, which targets the part of agent runs where context costs balloon fastest. The roadmap calls for additional MCP server types beyond compression once the marketplace launches.
Market data reflects a young subnet finding its footing. Price sits at 0.00930 TAO with a market cap around 9,132 TAO and a current emission share near 0.93%, smoothing to roughly 1.22% on the EMA that drives Taoflow. The seven-day window is up 21.7% on a positive 292 TAO net inflow, while the thirty-day window is down 14.9%, consistent with a subnet that ran up earlier and is now consolidating. The pool holds 2,380 TAO with root proportion near 0.47, meaning roughly half the pool depth is still protocol-sourced. Chain buy rate sits at 0.70%, which is low but, per the dTAO mechanics, not itself a risk.
Development is live. The GitHub repository at DendriteHQ/SOMA shows the last push on 2026-05-18, five contributors across 312 commits, and recent merges covering API rate limiting, validator settings, and sample task dumps. The team operates as Dendrite (DendriteHQ), with co-founder Oli posting regularly through @SomaSubnet and @oli_soma. The team has confirmed an in-person appearance at @proofoftalk in coordination with @DendriteHQ.
- Execution risk: SOMA depends on a centrally operated platform that orchestrates submissions, screening, and validator scoring. The on-chain weight setting is decentralized; the upstream pipeline that feeds those weights is not. If the platform stalls, the competition stalls.
- Competition: Anthropic's own MCP ecosystem and tool registries from large hosted providers are the obvious centralized alternatives. Inside Bittensor, multiple infrastructure subnets are vying for the agent stack from different angles.
- Concentration: A Gini coefficient of 0.68 across the top 100 positions points to moderately concentrated ownership or stake distribution. Larger positions could meaningfully move the pool.
- Liquidity: Pool depth of 2,380 TAO is on the thinner side of subnets with similar market caps. Larger entries and exits will move price.
- Roadmap dependency: The thesis assumes the marketplace launches in Q3 2026 with at least one production MCP. If that slips, the "winner-takes-all weekly" pattern stays a competition layer rather than a real service market.
Into the next one.
