The Agent

The Agent page is the heart of CCD: a persistent AI engineer that plans, writes, executes, and verifies work on your dedicated nodes.

Sessions

Each conversation with the agent is a session, shown as a tab. Sessions are persistent. You can leave, come back, and continue where you left off. Past sessions live in History.

Give the agent outcomes, not instructions: "Add authentication to the Flask app in ~/projects/shop and make the tests pass" works better than a series of small commands. The agent breaks the goal into steps, executes them on your node, and streams every file change and command as it happens.

Approvals

Agent tools are permission-scoped. Depending on the active mode, sensitive operations (executing shell commands, writing files, dispatching to other nodes) pause and ask for your approval before running. You'll see exactly what the agent wants to do and can approve or reject it. Read-only modes never modify anything.

Agent modes

Eight specialized modes tune what the agent focuses on and which tools it may use:

Mode Focus Access
👨‍💻 Developer (default) Full-stack development, production-ready code, tests Standard: exec requires approval
🔒 Cyber/Security Security audits, vulnerability scanning, threat analysis Security-scoped tools
📊 Research & Analysis Data analysis, research synthesis, documentation Read-only
🏗️ Architect System design, scalability planning, tech evaluation Elevated: exec & writes require approval
⚙️ DevOps CI/CD, infrastructure automation, deployments Elevated: infra tools require approval
📚 Teacher Explanations, tutorials, concept breakdowns Read-only
🐛 Debugger Root-cause analysis, log analysis, surgical fixes Standard: writes require approval
🔍 Code Reviewer Quality, security, and performance review Read-only

Switch modes from the Agent page. The mode governs both the agent's behavior and its tool permissions.

Run modes

A single toggle on the Agent page picks how much of your fleet a run uses. All four modes work in both the classic Agent and Agent HYPER.

Safety note: if a task is locked to a specific project folder, MULTI and COLLAB automatically collapse to a single node so parallel agents never fight over the same files. DOMO is reasoning only and does not write to your project.

How time-based modes are billed

MULTI, COLLAB, and DOMO rent multiple nodes at once, so they are metered by wall-clock time at the per-hour rate above, charged by the minute. A run needs a small credit buffer to start (a few minutes' worth), shows its running charge as it goes, and stops itself cleanly if your balance runs out. SINGLE runs stay token-based.

Bring your own AI keys (BYOK)

You can attach your own AI provider API keys (DeepSeek and Groq are routable today; keys for other providers can be stored for future use) from the API Keys manager on the Agent page. Keys are encrypted at rest. When you add a funded DeepSeek key, the agent uses your key first, so your runs keep going on your own quota. See API access for details.