Concepts
Projects and tasks
Kelvar organizes work in three layers. Home is your continuous conversation with the system. Projects are goals with their own context and permissions. Tasks are pieces of work you hand to an agent.
#Home
Home is one ongoing conversation. Type or talk to it. The Home coordinator acts as a security director for your whole installation. It can direct every project, create new agents and projects, and coordinate a response that involves several agents.
Home is where new interfaces, agents and projects are authored.
#Projects
A project is a goal with a policy. What an agent may do in one project says nothing about another, and an inner project can only narrow the permissions of the one around it. It can never widen them.
Setup creates five projects.
| Project | Focus |
|---|---|
| Cloud Config & IaC Reviewer | Infrastructure-as-code and configuration review in GitHub repositories and local checkouts. Live cloud scanning is a preview. |
| Zero-Click & CVE Response | Zero-click RCE and deserialization sinks, SSRF, Log4Shell-class CVEs and emergency virtual patching. |
| Code & Supply-Chain Audit | Lockfile and manifest auditing, GitHub Actions pinning and typosquatting. |
| DLP, Secrets & Compliance | Leaked keys and tokens, private keys, and evidence for SOC 2, HIPAA and PCI DSS. |
| 24/7 SOC & Threat Hunting | Alert triage, lateral movement detection and threat hunting. |
You choose which drivers and applications each project may use. Give a project the smallest set it needs. See Permissions and sandboxing.
#Tasks
A task is a run. You hand it to an agent in a project, then leave it running and come back later. Messages you send afterwards continue the same task.
To start one, choose a project, choose one of its agents, and write what you want. If the agent list is empty, create an agent first.
Tasks are durable. A task pauses when it is waiting for something, for example a child report or your approval, and wakes when that arrives. It does not hold a process open while it waits.
#Activity
Each task has an Activity trace that shows what the agent did, in the order it happened. It lists each tool call, the result, and any refusal. Activity is a view of the ledger for that run. See Ledger and marks.