brime 0.2.0

Official Python SDK for the Brime API (search · extract · research)

brime

The live-web toolkit for AI apps. One API key. One SDK. Search, scrape, and research the open web — clean output, sane defaults, no plumbing.

Documentation | Get Your API Key | Changelog | GitHub

pip install brime

Python 3.9+. Sync and async clients, exponential-backoff retry, with_raw_response observability, fully typed (py.typed). Single dependency tree: httpx + pydantic. Released via npm-style Trusted Publishing — every wheel ships with PyPI provenance.

Why brime?

• One key, three primitives. search, extract, research — the shape every AI app needs from the web.
• Tuned defaults. No depth selectors, no round counters, no knobs to babysit. The gateway is tuned for you; you pass a query, you get a clean answer.
• Drop-in compatible. Already on Tavily, Exa, or Parallel? Point their SDK at our adapter URL and your code keeps working. Migrate when you're ready.
• Honest pricing. Flat per-call credits. 0.5 for search, 1 per URL for extract, 5 for research. No surprises.

30 seconds

from brime import Brime

brime = Brime()  # reads BRIME_API_KEY

# Live answer + ranked sources, sub-second.
result = brime.search(query="what changed in the latest TypeScript release")
print(result.answer)

That's the whole shape. Same pattern for extract and research.

What you can build

Search the open web

result = brime.search(
    query="tesla earnings",
    topic="finance",       # optional: news / general / finance recency hint
    time_range="week",     # optional: day / week / month / year
    domains=["sec.gov"],   # optional allow-list
)

Turn any URL into clean markdown

result = brime.extract(urls=[
    "https://example.com",
    "https://en.wikipedia.org/wiki/BM25",
])

for r in result.results:
    print(r.url, len(r.markdown))
for f in result.failed:
    print("skipped", f.url, f.error.message)

Handles HTML, PDF, DOCX, and JavaScript-heavy SPAs. The smart-clean pipeline strips chrome, nav, cookie banners, and template noise — what's left is the article.

Multi-step research with citations

result = brime.research(
    query="compare frontier coding models with concrete benchmark numbers",
)

print(result.answer)
print(f"{len(result.sources)} sources cited")

One call, ~30–90 seconds, real synthesis from real sources.

Live progress? Stream it:

for evt in brime.research_stream(query="..."):
    print(evt.event, evt.data)
    if evt.event in ("complete", "error", "timeout"):
        break

Authentication

export BRIME_API_KEY="sk-brime-..."

Brime()                        # uses BRIME_API_KEY
Brime(api_key="sk-brime-...")  # explicit
Brime(base_url="https://...")  # staging override (or BRIME_BASE_URL env)

Get a key at brime.dev — the free tier comes with 1,000 credits/month and no card.

Async

Every method is mirrored on AsyncBrime:

import asyncio
from brime import AsyncBrime

async def main():
    async with AsyncBrime() as brime:
        result = await brime.search(query="python async io")
        print(result.answer)

asyncio.run(main())

Async streaming works the same way:

async with AsyncBrime() as brime:
    async for evt in brime.research_stream(query="..."):
        if evt.event == "complete":
            print(evt.data)
            break

Error handling

Typed exceptions, predictable surface area:

from brime import (
    BrimeError,
    AuthenticationError,
    RateLimitError,
    InsufficientCreditsError,
)

try:
    brime.search(query="...")
except AuthenticationError:
    ...  # bad key
except RateLimitError:
    ...  # back off
except InsufficientCreditsError:
    ...  # top up
except BrimeError as e:
    print(e.code, e.message)

Idempotency, baked in

extract calls require an Idempotency-Key — the SDK auto-generates one per call so accidental retries never double-charge. Pin it yourself for cross-process dedup:

brime.extract(
    urls=["https://x"],
    idempotency_key="user-42-prefetch-2026-05",
)

Configuration

Constructor arg	Env var	Default
api_key	BRIME_API_KEY	— (required)
base_url	BRIME_BASE_URL	https://api.brime.dev
timeout	—	30.0 seconds

Per-call override: brime.search(query="...", timeout=60).

Already using Tavily, Exa, or Parallel?

You don't have to rip them out. Brime exposes wire-compatible adapters:

TavilyClient(api_key, api_base_url="https://api.brime.dev/tavily")
Exa(api_key=..., base_url="https://api.brime.dev/exa")
Parallel(api_key, base_url="https://api.brime.dev/parallel")

Same response shapes, same code. Switch to the native brime SDK when you want the extras (research synthesis, SSE streaming, smart-clean extract).

Links

• Docs — docs.brime.dev
• API reference — docs.brime.dev/api-reference
• Status — brime.dev
• Issues — github.com/brime-dev/brime-python/issues

License

MIT © Brime