Concepts
Security agents
An agent in Kelvar is a role plus a ceiling. The role says what the agent is for. The ceiling says the most it may ever touch. Kelvar provisions seven agents during setup, and you can create more.
#The seven agents
| Agent | Focus | Works in |
|---|---|---|
| Zero-Click Rapid Defender | Traces untrusted input to dangerous sinks, hunts SSRF and critical CVEs, and drafts virtual patches. | Read-only rooms |
| GitHub & Supply-Chain Auditor | Audits repositories, pull requests, workflows and dependency manifests for CVEs and supply-chain risk. | Read-only rooms |
| Cloud Config & IaC Reviewer | Reviews Terraform, CloudFormation, Dockerfiles and Kubernetes manifests for over-privileged IAM and public exposure. | Read-only rooms |
| DLP, Secret Leak & Credential Hunter | Scans code, git history, CI logs and container layers for secrets, keys and personal data. | Read-only rooms |
| Endpoint, Browser & Phishing Forensic Guard | Opens suspicious links and OAuth consent pages in an isolated browser, and audits host persistence and connected Linux servers through a read-only collector. | Isolated browser and process rooms |
| 24/7 SOC & Threat Sentinel | Investigates SIEM and webhook alerts, correlates signals, escalates confirmed incidents and keeps compliance evidence tied to the ledger. | Read-only rooms |
| Autonomous Security Patch & Hardening Engineer | Authors, tests and verifies minimal patches, with a reversible change journal. | Isolated git worktrees |
#What an agent is made of
Underneath, an agent is two JSON files in your registry, not a program.
- A policy file that sets the ceiling, which lists the tools and room kinds the agent may ever use.
- A role artifact that holds the instructions the agent follows.
Because an agent is data, you can have as many as you want, and Kelvar can inspect and replay any of them. When you create an agent in the app you choose a project, a model and a reasoning level, then review its capabilities before it is saved.
#Rooms
A room is the sandbox where an agent's tools act. Kelvar ships three kinds.
| Room | What it allows |
|---|---|
| Workspace | Reading, listing and searching files, and writing files inside admitted paths. |
| Process session | Running commands in a sandboxed shell. |
| Git worktree | The same file tools, in an isolated worktree on its own branch, so a change never touches your working copy. |
An agent can only use a tool that its ceiling allows and that its room provides. If the two disagree, setup fails with a clear error rather than guessing.
#Choosing a model per agent
Each agent has its own model and reasoning level. A common setup is a stronger model for the Zero-Click Rapid Defender and the Patch Engineer, and a smaller one for routine alert triage. You can use any provider that Kelvar supports. See Bring your own model.
#Delegation
When an agent is asked to audit a path outside its room, it does not ask for broader access. It authors a child worker with a read-only room on that path, starts a child run, waits for the verified report, and then summarizes it. The child's access ends with its run.