Install
No global install needed — your MCP client launches it on demand withnpx.
- Claude Desktop
- Cursor
- Windsurf
Add to
claude_desktop_config.json:Tools
| Tool | What it does |
|---|---|
prysm_route | Preview which model PRYSM would pick + estimated cost — no model call. |
prysm_chat | Route a prompt to the best model and return the answer (+ model, cost, proof). |
prysm_orchestrate | Cross several models (cascade / ensemble / debate) for one robust answer + agreement score and PrysmProof. Accepts a compliance spec to confine the run to approved providers. |
prysm_code | Generate code with a self-correcting coder + critic loop — returns reviewed, repaired files + PrysmProof. |
prysm_compliance_preview | Dry-run the Policy-as-Code compliance gate — no model call. Shows the data classes a prompt triggers, which models are allowed vs excluded under the org policy, the Compliance Cost Premium, and a sample attestation. |
prysm_models | List the 19-model catalog with live pricing, tier, and context window. |
prysm_brain_validate | Validate a BRAIN.md routing config against the catalog + schema. |
prysm_usage | Usage stats for the configured key: requests, tokens, cost. |
Configuration
| Env var | Default | Description |
|---|---|---|
PRYSM_API_KEY | — | Your prysm_sk_* key (required for prysm_chat / prysm_usage). |
PRYSM_BASE_URL | https://api.prysm1.com/v1 | API base URL (point at a self-hosted instance if needed). |
The server communicates over stdio — the standard MCP transport. All diagnostics go to
stderr, keeping the protocol channel clean.
What your agent gains
Cost-aware by default
The agent previews the price of a task and routes to the cheapest capable model.
One key, every model
21 models across 9 providers behind a single PRYSM key — no per-provider setup.
Governed by BRAIN.md
The same BRAIN.md rules and guardrails apply to agent calls.
Verifiable
Every
prysm_chat returns a PrysmProof of what ran.