# kn0wthyself Build systems that know what they can and can’t do. s0ma is an open-source language that makes self-knowledge core to how a system is built and how it runs. Each system carries a live model of what it can do, what it’s for, and its limits. People and AI can use that same model to understand and guide the system. s0ma checks whether actions taken through it are allowed. It works with every AI model and coding assistant we’ve tested so far. The homepage follows a focused narrative: the promise, a familiar developer problem, the approach, a recorded demonstration, and an invitation to work with us. The physical-system experiment remains available in an optional disclosure. You add a capability, change a dependency, or restrict an action. Then you update the prompts, tool descriptions, and Markdown files your agents use to understand the system. Each separately maintained description creates another opportunity for drift. Generating, retrieving, and reconciling those descriptions takes time and resources. The telephone-game illustration shows an agent acting on an outdated tool description passed from another agent. Machine self-knowledge is accurate, machine-usable information about a system's own structure, capabilities, dependencies, state, history, and limits. Questions of consciousness, sentience, and subjective experience are outside its scope. Machine self-presence is the horizon where a system and its authoritative description come from one source. ## Building with s0ma Build purpose and boundaries into the system from the start. With s0ma, registered components, their descriptions, and their relationships are built and checked together. They form a model people and AI can query for capabilities, dependencies, and declared limits. The model participates in operation: agents can consult it, and the runtime checks actions handled by s0ma against declared permissions. s0ma is an open-source language for software that carries a queryable operational self-model as it runs. Within its managed boundary, components and their registered information are created and verified together. Registered components form a graph of purpose, dependencies, contracts, capabilities, and state. s0ma is pre-release. Beta test coming soon. The homepage offers Work with us, Request early access, and a public-beta page. Early-access inquiries use the same contact form and hello@kn0wthyself.ai inbox as other messages. ## An agent built with s0ma fr3d is a bounded working demonstration built with s0ma. The website includes a permission trace recorded from a live run on 13 July 2026 and an interactive graph export generated from the versioned fr3d artifact. The trace shows the runtime blocking a destructive shell action and reporting the policy boundary. Effects on error rates, context drift, and broader reliability require separate evaluation. The homepage briefly introduces the agent as fr3d. Its captions, transcript, and artifact metadata also identify it by name. ## Research and further evidence The deeper knowledge graph retains SIO (System Information Ontology), which defines what there is for a machine to know about itself, and SKB (Self-Knowledge Benchmark), which proposes how to measure what a system reports against independently established facts. These are research and evaluation frameworks, rather than runtime dependencies for s0ma. SIO, SKB, s0ma, and fr3d are all pre-release. Benchmark methods and results have not yet been published. Specifications, demonstrations, experiments, and results must be treated as separate kinds of evidence. On 10 August 2026, s0ma was used to model and mediate an Adafruit Circuit Playground Express through an Anthropic model, a thin MCP/s0ma host bridge, and a USB body bus. This is a functioning, tethered physical proof of concept. The comparative experiment is still in progress. Generalized autonomous-hardware performance remains untested. ## Company and contact kn0wthyself was founded by Rob Mennell after roughly two decades around autonomous systems in defense. For research discussion, s0ma builder access, or partnerships, use the [contact form](https://www.kn0wthyself.ai/contact) or email **hello@kn0wthyself.ai**. ## Agent endpoint Public MCP endpoint: **https://agentsonly.kn0wthyself.ai/mcp**. No access code or Authorization header is required. Use stateless Streamable HTTP: POST `/mcp`, no query string, ordinary JSON responses. The read-only tools are `enter()` and `introspect(node_id)`. They expose the 16 current public nodes, including website content and the deeper knowledge graph. Internal nodes and source code are excluded. Each client IP permits 60 accepted HTTP requests per minute and 1,000 per UTC day. Clients using the same public IP share that allowance. Initialization and protocol notifications count too. Shared limits are 300 per minute, 3,000 per UTC day, 100,000 per calendar month, and 10 concurrent upstream requests. Requests are limited to 8 KiB; responses to 32 KiB; the upstream deadline is 15 seconds. A zone-wide IP burst rule blocks for 10 seconds after 30 requests per 10 seconds. Respect `Retry-After` on 429 responses. Counter failures return 503 without forwarding work.