What runs underneath
Every request to eutonomous passes through malgra, a gateway we wrote in Rust. It is where the security features on our home page actually happen. You do not need to know any of this to use the service — this page is for anyone who wants to see how it works.
On every request
In order, for every request, in either API format. Each step is configuration, so a team's security profile decides which ones apply.
Requests arrive in the OpenAI or the Anthropic API format and are translated as needed, so Claude Code and an OpenAI-SDK script hit the same endpoint with no client changes.
The key identifies the caller and resolves the security profile its team assigned to it. A key the gateway does not recognise is refused, never served unprotected.
Platform rules for known attack patterns, then your own. Depending on the profile a match blocks the request or is recorded and let through.
Tool output inside the conversation is treated as data. Instructions hidden in it are flagged — the same words typed by you are not.
Emails, phone numbers, IBANs, card numbers, IP addresses and API keys become placeholders before anything leaves for the model provider.
Only to the EU endpoints on the sovereignty ledger. The provider sees the prompt with placeholders, never the original values.
On the way back the placeholders are swapped for the real values, so your agent gets an answer it can use as is.
Every security decision is appended to a hash-chained record signed with the instance key — what was decided and why, never the prompt.
Built to be checked
Download your account's record and check every signature and the chain offline. A changed or missing entry breaks it.
Every protection is explicit configuration, per profile. What a team has switched on is what it gets — no silent defaults.
malgra runs eutonomous in production today, and we intend to release it as open source so anyone can read and run it.
Related projects
Three separate projects that work well with malgra. None is part of eutonomous, and you need none of them to use it.
Secure MCP: the MCP tool model with authentication, per-message encryption and agent-to-agent coordination, so tools can run safely over a network. malgra speaks it natively.
A terminal coding agent that can prove to malgra how it is sandboxed and which text came from tools — the first client to use malgra's attested features.
Encrypted long-term memory for agents: notes that link themselves, personal data kept out of the index, and embeddings that can go through malgra so a workspace holds no credentials.