The prysm command ships with the Python SDK. Install it once and you
have routing previews, the model catalog, BRAIN.md validation, usage stats, one-shot
chat, multi-model orchestration,
self-correcting code generation, and
compliance previews from the terminal.
Configure it with the standard environment variables:
export PRYSM_API_KEY="prysm_sk_your_key"
# Optional:
export PRYSM_BASE_URL="https://api.prysm1.com/v1"
Commands
| Command | What it does |
|---|
prysm route "<prompt>" | Preview routing + estimated cost (no model call). |
prysm models | List the 19-model catalog with pricing, sorted by input price. |
prysm brain validate [path] | Validate a BRAIN.md (defaults to ./BRAIN.md). |
prysm usage | Usage stats for your key. |
prysm chat "<prompt>" | One-shot completion (routed) — prints the answer. |
prysm orchestrate "<prompt>" | Multi-model orchestration (v2) — prints the answer + which models ran. |
prysm code "<task>" | Code Mode (v2) — self-correcting code generation; prints the files. |
prysm compliance preview "<prompt>" | Dry-run the compliance gate — classes, allowed/excluded models, CCP (no model call). |
prysm --version | Print the installed version. |
prysm route
Preview which model PRYSM would pick — and what it would cost — without spending
anything. Add --mode to force quality, balanced, or agility.
$ prysm route "write a python function to sort a list"
model: deepseek-v3.2 (DeepSeek V3.2)
mode: balanced tier: budget
why: Code: 95% cheaper than GPT-5.2
est cost: $0.000700 / 1K tokens (in 0.28 / out 0.42 per MTok)
signals: code
prysm route "summarize this 30-page report in depth" --mode quality
prysm models
$ prysm models
MODEL TIER IN $/MTok OUT $/MTok
mistral-nemo budget 0.02 0.02
gpt-5-nano budget 0.05 0.4
gemini-2.5-flash-lite budget 0.1 0.4
...
21 models
prysm brain validate
Validates ./BRAIN.md by default, or a path you pass. Exits non-zero when the config is
invalid — handy in CI.
$ prysm brain validate
VALID (BRAIN.md)
$ prysm brain validate config/BRAIN.md
INVALID (config/BRAIN.md)
error: unknown model in rule: ghost-model
Wire it into a pre-commit hook or CI step so a broken BRAIN.md never ships:prysm brain validate || exit 1
prysm usage
$ prysm usage
user: usr_demo plan: free
requests: 128
tokens: 512340
cost: $0.214500
prysm chat
A one-shot routed completion. The answer prints to stdout; the routing summary prints to
stderr, so you can pipe the answer cleanly.
$ prysm chat "explain a bloom filter in two sentences"
A Bloom filter is a space-efficient probabilistic data structure...
[routed to DeepSeek V3.2 · $0.000058]
# Pipe just the answer (routing note goes to stderr):
prysm chat "name three uses for a heap" > answer.txt
prysm orchestrate
Run a prompt across several models and synthesize one answer — then see which models
contributed and how strongly they agreed. Pick the objective with --policy; force an
execution shape with --strategy. See How orchestration works.
$ prysm orchestrate "Compare three database designs for a 40-person team" --policy depth
Across the three designs, the trade-offs cluster around consistency vs. operational cost...
policy: depth strategy: ensemble_moa
models: claude-sonnet-4.5, gpt-5.2, gemini-3.1-pro
agree: 0.81 confidence: 0.88
cost: $0.004812 (est) proof: sha256:a1b2c3d4e5f60718
| Flag | What it does |
|---|
--policy | efficiency, balanced (default), or depth. |
--strategy | Force single, cascade, ensemble_moa, rank_fuse, decompose_and_route, self_consistency, or debate. |
--k | Ensemble / sample width. |
--max-cost | Soft budget hint, in USD. |
prysm code
Generate code with a self-correcting, multi-model loop: a coder writes it, a separate
critic reviews it, and the coder repairs against the feedback. The files print to stdout;
the run summary (verdict, models, cost, proof) prints to stderr, so you can pipe the code
cleanly. See How Code Mode works.
$ prysm code "Write a Python function to add two numbers" --policy balanced
===== main.py (python) =====
def add(a, b):
"""Return the sum of two numbers."""
return a + b
[PASS] policy: balanced language: python iterations: 1
coders: qwen-2.5-72b reviewer: deepseek-r1
cost: $0.000412 (est) proof: sha256:7f3a9c2b1e4d8a06
# Write generated files into ./out (creates directories as needed):
prysm code "Implement a CLI todo app in Python" --policy depth --out ./out
# Single shot, no critic loop:
prysm code "Write a Python hello world" --no-review
| Flag | What it does |
|---|
--policy | efficiency, balanced (default), or depth. |
--language / --lang | Force the target language (auto-detected if omitted). |
--coder-model | Force the model that writes the code. |
--reviewer-model | Force the critic model. |
--max-iters | Generate + up to max-iters - 1 repairs. |
--no-review | Single shot — skip the critic loop. |
--max-cost | Soft budget hint, in USD (stops the repair loop). |
--out | Write files to this directory instead of stdout. |
prysm code exits 0 when the review passes and 1 when it fails (or on error), so it
drops cleanly into CI and scripts.
prysm compliance preview
Dry-run the Policy-as-Code compliance gate against a prompt — no
model is called. It reports which sensitive data classes the text triggers, which models
the policy allows vs excludes, and the Compliance Cost Premium. Build the policy from flags
(and/or a JSON --policy-file; flags override the file).
$ prysm compliance preview "wire payment to DE89370400440532013000" \
--jurisdiction EU --frameworks GDPR --data-residency EU
classes: FINANCIAL
status: OK
allowed: 2 — mistral-nemo, mistral-medium-3
excluded: 17
CCP: +$0.000000/MTok (0.0%) [mistral-nemo vs unrestricted mistral-nemo]
attest: sha256:a1b2c3d4e5f60718
SECRETS blocked is a hard stop:
$ prysm compliance preview "deploy key AKIAIOSFODNN7EXAMPLE" --block-classes SECRETS
classes: SECRETS
status: BLOCKED — no approved model can serve this request
allowed: 0 — (none)
excluded: 19
| Flag | What it does |
|---|
--allow | Comma-separated provider allowlist (e.g. anthropic,openai). |
--deny | Comma-separated provider denylist. |
--jurisdiction | Allowed jurisdictions (e.g. US,EU). |
--frameworks | Required frameworks (e.g. GDPR,CCPA). |
--certifications | Required certifications (e.g. SOC2,HIPAA). |
--data-residency | Required data regions (e.g. EU). |
--block-classes | Data classes to block: PII,PHI,FINANCIAL,SECRETS. |
--zero-retention | Require a zero-retention provider mode. |
--policy-file | JSON Policy-as-Code spec; inline flags override it. |
prysm compliance preview exits 0 when at least one approved model can serve the
request and 1 when the policy blocks every model — so it gates CI cleanly.
Exit codes
| Code | Meaning |
|---|
0 | Success (and, for brain validate, the config is valid). |
1 | A runtime error, or an invalid BRAIN.md. |
2 | A file couldn’t be read (e.g. missing BRAIN.md path). |
