pareta

Python client for Pareta. One model id — "auto" — and Pareta plans each request, routes it to benchmark-proven open specialists, verifies the result, and falls back to a frontier model when that's the right call. One request, one bill; you never pay for Pareta's orchestration or cold starts.

pip install pareta        # or: uv add pareta / poetry add pareta

from pareta import Pareta

pa = Pareta.from_env()                       # reads PARETA_API_KEY
# or: Pareta(api_key="pareta_sk_…", base_url="https://api.pareta.ai")

resp = pa.chat.completions.create(
    model="auto",                            # the routing brain — the product
    messages=[{"role": "user", "content": "Extract the total from this invoice: …"}],
)
print(resp.choices[0].message.content)

# Streaming (progress while Pareta plans + executes, then tokens)
for chunk in pa.chat.completions.create(model="auto", messages=[...], stream=True):
    print(chunk.choices[0].delta.content or "", end="")

Async mirrors the sync client:

from pareta import AsyncPareta

async with AsyncPareta.from_env() as pa:
    resp = await pa.chat.completions.create(model="auto", messages=[...])

Is it actually good? Measure it on YOUR data

Don't take the routing brain on faith — benchmark it. pa.evals runs "auto" head-to-head against frontier models on your own ground truth and prices every contender honestly:

run = pa.evals.runs.create(eval_set=es.id, models=["auto"],
                           frontier=["claude-opus-4-7"], wait=True)

And watch what your live traffic is doing — spend, success rate, and the projected savings vs calling a frontier directly:

pa.auto.metrics()          # requests, success rate, spend, savings vs frontier
pa.auto.compare_frontier(  # one prompt, metered, side-by-side with auto
    model="gpt-5.5",
    messages=[{"role": "user", "content": "…"}],
)

Inference is OpenAI-compatible

You don't even need this SDK to call Pareta — point the openai client at base_url + your key and set model="auto":

from openai import OpenAI
client = OpenAI(api_key="pareta_sk_…", base_url="https://api.pareta.ai/v1")
resp = client.chat.completions.create(model="auto", messages=[...])

This SDK's unique value is everything AROUND that call — evals on your data, auto metrics, and the benchmark catalog — as Python methods, a CLI, and an MCP server.

Discovery

pa.tasks.match resolves a plain-language intent to the benchmarked task (or capability lane) Pareta covers it with — feed the matched task into pa.evals to prove "auto" on your own data:

m = pa.tasks.match("extract the key fields from these contracts")
m.type, m.chosen.task_id                     # "task", "contract-key-fields"

for model in pa.models.list():               # everything your org can call
    print(model.id)

Auth

Mint a pareta_sk_ key in the dashboard (key management is browser-only) and pass it as api_key= or via PARETA_API_KEY. The SDK only ever consumes a key; it never creates, lists, or revokes them.

CLI

pip install "pareta[cli]" adds the pareta command (or pipx install "pareta[cli]" for an isolated, always-on-PATH install):

export PARETA_API_KEY=pareta_sk_…

pareta chat "Summarize this contract: …"               # model:"auto" by default
pareta auto metrics                                     # your auto traffic, rolled up
pareta auto compare "…prompt…" --frontier gpt-5.5       # auto vs a frontier, metered

pareta tasks match "extract fields from invoices"       # intent → task
pareta evals run --prompt "extract the total from each invoice" \
  --file rows.jsonl --models auto --frontier --wait     # prove auto on your data

Add --json to any command for machine-readable output; pareta --help (or pareta <group> --help) documents the full tree — chat, auto, tasks, models, evals, audio.

MCP server

pareta-mcp is a Model Context Protocol server (stdio) that exposes Pareta to an AI agent (Claude Desktop, Cursor, …) as tools — chat (defaults to model="auto"), auto_metrics, compare_frontier, run_eval / get_eval_run, discovery (match_task, list_tasks, get_task, list_models), and audio (transcribe, speak).

Run it in its own isolated environment — like any MCP server it has its own dependency tree, so don't pip install it into an app/project venv. The simplest is uvx (no install, runs on demand). Register it (Claude Desktop → Settings → Developer → Edit Config):

{
  "mcpServers": {
    "pareta": {
      "command": "uvx",
      "args": ["--from", "pareta[mcp]", "pareta-mcp"],
      "env": { "PARETA_API_KEY": "pareta_sk_…" }
    }
  }
}

Prefer a persistent install? pipx install "pareta[mcp]" puts pareta-mcp on your PATH in a dedicated venv — then use "command": "pareta-mcp". (Avoid a plain pip install "pareta[mcp]" into a shared environment: its mcp/starlette dependencies can clash with an app's FastAPI, and the console script may not land on your PATH.) Inference, eval, and audio tools spend money; your MCP client's per-tool-call approval is the guardrail.

Errors

All errors subclass pareta.ParetaError:

Idempotent GETs and 429/5xx/timeouts are retried with exponential backoff (max_retries, default 2).

Status

Live: model="auto" inference (buffered + streaming) with pa.auto metrics and frontier comparison, plus models, tasks (browse + match), evals (bring-your-own-data, with "auto" as a first-class contender), and audio — and two interfaces over it: the pareta CLI (pip install "pareta[cli]") and the pareta-mcp MCP server (pip install "pareta[mcp]"). Sync + async clients.