Coding agents

PrismaLens doesn’t include a model or an agent. It drives a coding agent on your machine over the Agent Client Protocol (ACP), the way an editor does, and that agent does the reading and the reasoning.

Supported agents

OpenCode, Claude Code, Codex, Gemini CLI and deepagents work out of the box. Other ACP agents can be added from the ACP registry, or as a local command, in Settings → Agent → Add agent.

Which one runs

1. PRISMALENS_HARNESS, if set when pl up starts, pins one, for example PRISMALENS_HARNESS=claude-code pl up.
2. Otherwise, the agent chosen in Settings → Agent.
3. Otherwise, the first agent that can run, in the order listed above.

Signing in

The agent uses its own sign-in, as it does in your terminal. You can also enter an API key in Add agent. PrismaLens stores that key encrypted and never shows it again. It never reads or stores the agent’s own login.

What the agent may do

You pick one of four permission levels. Each selects the agent’s own mode and decides which of its requests wait for you:

Level	What it means
Ask always (default)	Reads and searches run; every other request waits for your approval.
Auto-accept edits	Reads and file changes are approved; every other request waits for you.
Auto	Runs the agent’s own review mode where it has one; requests are approved.
Full access	Every request is approved. The agent can do anything your user can.

A run you start uses Ask always unless Settings or the agent’s own configuration sets another level, so check both before you rely on approvals. A run started by an alert has nobody to ask. It runs at Auto and refuses anything the agent asks. PrismaLens adds no sandbox. See What PrismaLens reads.