Agentome GitHub
Contents

Security

This page is for whoever answers for the company's security: what an agent can reach, where each secret lives, and what Agentome does not protect. It describes an installation on the company's server with places under gVisor, as on Deploy on your server; a person's computer gets less, as the last section says.

An agent is a program that runs commands

An agent reads tasks, issues, code and documents that other people wrote, and any of them can carry text that tells it to do something else. So Agentome does not rely on the agent behaving: each run gets only what its task needs, and the rest is out of its reach whatever it does.

Where a run happens

A person's agents run in their place, a container the platform makes for them on the server. Each run in it gets two containers of its own, made when it starts and deleted when it ends:

So a command the agent was talked into cannot read the subscription's token, and the CLI never runs a command itself. Every place and every run has a kernel of its own under gVisor, not the host's. A place has a read-only root, no Linux capabilities, at most 2 CPUs and 8 GB of memory, and no DNS for names outside. The service that makes containers holds no Docker socket: a guard does, with no network, and passes only the calls that make places and runs. Places, the sandbox and a run's own Docker never get the host's folders, ports or namespaces.

A place runs a person's subscription only while that person is at work in Agentome, and the person signs the CLI in themselves, in the place's terminal. Agentome neither reads nor keeps the sign-in, and never runs a subscription for anyone else.

What a run can reach

A place has no way out but the egress proxy, which knows every call's token and lets it through to its own list only:

The proxy takes only HTTPS through CONNECT, to the name that was asked for, on port 443 unless the list names another. It checks the address a name resolves to, never connects to the server's own machine or to link-local addresses, and reaches the company's private networks only for a host listed by its full name. It reads the server name of each connection and nothing inside it. Every host it lets through or refuses is in the run's log, and a refused one shows in the project's Network section with Allow beside it.

Where the secrets are

Secret Where it is Who uses it
A person's subscription sign-in Their place's home volume, on the server The CLI in a run's CLI container
The code host, Jira and MCP server tokens The Agentome database, encrypted The server's gateway and git proxy
Model API keys, for agents on the server The Agentome database, encrypted The server's relay, which adds them to a call
A run's token The run, until it ends The run's tools, git and way out
The installation's secret key The data folder, or the environment on Kubernetes The server

Integration tokens never reach a place or a computer. An agent's tools go through the server's MCP gateway, with the run's token, which applies the agent's policies to each call: allow, ask a person first, or deny. Its git goes through the server's git proxy, which adds the code host's token on the way. The model relay does the same for API keys. A run's token stops working when the run ends.

The database keeps these secrets under AES-256-GCM, with a key drawn from the installation's secret key. A backup holds the database and not the key. agentome rotate-key seals everything under a new key when the old one may have leaked.

What an agent can push

The git proxy opens only the repository of the run's task, and holds every push to rules the agent cannot change:

A change request is opened by a tool, under the agent's policies, and merged by a person.

People and access

What Agentome does not protect

Found something? Write to the address in SECURITY.md; please do not open a public issue.

Edit this page on GitHub