When Mira Murati, former CTO of OpenAI, quietly launched her new venture Thinking Machines Lab last year, the crypto-native crowd paid attention—not because of AI, but because of the promise of decentralized intelligence. Today, her team released Inkling, an open-source model that claims to be the “best Western open-source model” for MCP (Model Context Protocol) scores. Listed on OpenRouter, Inkling is already live for developers. But as I dug into the technical signals, a familiar pattern emerged: bull market enthusiasm meets code-audit skepticism.
### The Context: Why Should Blockchain Care? Murati’s departure from OpenAI was accompanied by a public letter calling for responsible AI development. That message resonated with the Web3 ethos, where trustless systems overlap with ethical technology. Inkling is positioned as a model optimized for agentic workflows—tools that can call other tools, manage context, and coordinate across decentralized networks. This is exactly what a DAO needs: an intelligent agent that can propose, execute, and arbitrate governance actions without central gatekeepers.
The model’s key differentiator is its performance on MCP, a protocol that enables models to interact with external systems. In theory, this could power on-chain agents that sniff vulnerabilities, automate liquidity rebalancing, or even moderate community forums. But a quick glance at the available data raises red flags. No standard benchmarks (MMLU, HumanEval, GSM8K) are provided. The only metric is “impressive MCP scores”—a non-standard test that may measure engineering cleverness more than genuine reasoning.
### Core Insight: The Devil Is in the (Missing) Details Based on my experience auditing DeFi protocols, I’ve learned that every “best-in-class” claim must be stress-tested. Inkling is no different. The analysis I conducted on the announcement reveals a scarcity of verifiable data. The model’s architecture, training compute, and dataset are undisclosed. The “open-source” label is ambiguous—no license terms are specified. If it’s based on Llama or Mistral (common open-source bases), then Inkling might be a fine-tuned variant with a marketing spin.
The MCP emphasis is both a strength and a weakness. On one hand, it signals a deep focus on agent tool use—a field that is still nascent but critical for autonomous systems. On the other hand, such specialization often comes at the cost of general intelligence. In the words of the analyst, “an agent that can call APIs but cannot reason through a math problem may be a dangerous tool in the wrong smart contract.” Without seeing SWE-bench or GAIA scores, we cannot assess whether Inkling can handle the complex, multi-step tasks that DeFi automation demands.
Yet, there is a hidden opportunity here. The MCP protocol itself could become a standard for agent-to-agent communication in Web3. If Thinking Machines Lab open-sources not just the model but the protocol, it might bootstrap a new infrastructure layer for decentralized orchestrators. This is reminiscent of how Ethereum’s ERC-20 standard unlocked a wave of tokenization. We should watch whether MCP gets adopted by frameworks like LangChain or integrated into wallet architectures.
### Contrarian Angle: The Pragmatic Test Let’s be honest: the “best Western open-source” claim is a narrative play. It deliberately ignores models from the East (DeepSeek, Qwen) that may be more capable. The title is designed to appeal to Western developers who fear losing AI sovereignty. But does that make Inkling valuable? Possibly, if the team executes on a clear product vision.
From a blockchain perspective, the bigger risk is that Inkling is just another model in a sea of open-source releases. The real innovation might be in the business model: using OpenRouter as a dissemination channel suggests Thinking Machines Lab is prioritizing developer reach over direct monetization. This could be a “governance by attention” play—gain traction, then introduce a token or DAO structure to capture value. In my Prague Consensus Workshop, I saw how open-source communities can align incentives through tokenized voting. If Murati’s team adopts such a model, Inkling could become the first truly community-governed AI.
However, the lack of transparency in the initial release is a red flag. In the crypto world, we know that “trust us, we’re the best” often precedes a rug pull. Building for humans means building with verifiable proofs. The team should release a technical report, benchmark results, and a clear open-source license before asking developers to build on top of it.
### Takeaway: A Vision Beyond the Model Education is the ultimate yield. The buzz around Inkling is a perfect moment to remind ourselves that technology is only as valuable as the community that governs it. Whether Inkling becomes the backbone of decentralized agents or just another footnote depends not on its MCP scores, but on whether its creators embrace the principles they once championed: transparency, inclusion, and decentralized control.
I will be testing Inkling on a simple DAO voting agent this week. If it can handle a multi-step governance proposal without hallucinating, it might earn a place in my toolkit. Build for humans, not just nodes. The future of AI and Web3 is not about the smartest model—it's about the most trustworthy one.