Skip to content
PrismaLens Docs

Bring your own agent & key

PrismaLens ships no model and no bundled agent. It rents a coding-agent harness to do the read-only investigation (Tier-2), and calls a model you configure to write the report (Tier-1). Both are bring-your-own-key: no PrismaLens account, no subscription, no credits. You point it at whichever provider you already pay for, and PrismaLens reads the credential fresh from your environment on every run — it never stores a key (ADR-0006).

That’s the trade PrismaLens makes deliberately: instead of embedding a model and metering you for it, it stays local-first and Apache-2.0, and rides on credentials you already have.

The agent lineup (honest status)

Section titled “The agent lineup (honest status)”

PrismaLens drives the harness through one permission dial and reports what actually happened; the harnesses themselves differ in maturity. Here’s the honest state for the current release (pl --version):

  • claude-code — shipped, and the default. Runs sandboxed on Anthropic’s native Claude Agent SDK. Bring an Anthropic subscription (a signed-in Claude Code session) or an ANTHROPIC_API_KEY.
  • deepagents — shipped, the alternate lane. Driven over the Agent Client Protocol (ACP), it speaks to any OpenAI-protocol provider — Ollama, OpenAI, Groq, or any OpenAI-compatible endpoint.
  • codex — planned, not yet available. Reserved as a harness value, but selecting it exits with an error today. Don’t plan a current setup around it.

The full compatibility matrix — binaries, install commands, read-only fidelity, and every provider credential — lives in Providers & harnesses. It’s the single source of truth; this page is just the framing.

Model quality still matters

Section titled “Model quality still matters”

BYO-model doesn’t mean any model. An investigation drives an agent through many tool calls and then reasons over the result, so a small local model produces noticeably weaker root-cause analysis — and you’ll blame the tool, not the model. There’s a floor worth respecting; the recommended model classes per harness are in Providers → Picking a model that’s actually up to the job.

Next steps

Section titled “Next steps”