Skip to content

integrations.md

Latest commit

 

History

History
71 lines (56 loc) · 3.51 KB

integrations.md

File metadata and controls

71 lines (56 loc) · 3.51 KB
Raw

Integrations

tare integrates in-process (Python), via the local proxy (any language), or as an MCP server. This page is the full adapter reference; the README has the short version.

Python — tare-compress

pip install tare-compress
import tare

skeleton = tare.skeletonize(open("big.rs").read(), "big.rs")
slim     = tare.compact_html(raw_html)
out      = tare.compress(blocks_json, task="fix the login bug")

All functions: compress, skeletonize, compact_lossy, compact_html, compact_csv, slim_schema, telegraphic, deref_images, crush, expand. The wheel is abi3 — one build per platform, works on Python 3.9+.

Framework adapters

from tare.integrations import (
    compress_messages,
    LiteLLMHandler,
    CompressionMiddleware,
    langchain_chat_model,
    agno_model,
    strands_model,
    anthropic_with_tare,
    openai_with_tare,
)

All imports are lazy — the module loads cleanly even when the underlying framework is not installed.

Export What it does Where compression runs
compress_messages(messages, task="") Core helper — converts OpenAI-style messages to tare blocks, compresses, and returns a condensed list[dict] in-process (Rust binding)
LiteLLMHandler(task="") CustomLogger-compatible callback for litellm; mutates messages in-place before each sync or async call. Register with litellm.callbacks = [LiteLLMHandler()] in-process
CompressionMiddleware(app, task="") ASGI middleware (Starlette / FastAPI / raw ASGI) — intercepts JSON chat bodies containing messages and compresses them transparently in-process
langchain_chat_model(base) Subclass factory — returns a tare-compressing subclass of any LangChain BaseChatModel with a tare_task field in-process
agno_model(base) Subclass factory — returns a tare-compressing subclass of any Agno Model in-process
strands_model(base) Subclass factory — returns a tare-compressing subclass of any Strands Model (Bedrock-style messages) in-process
anthropic_with_tare(client_kwargs=None, base_url="http://127.0.0.1:8787") Returns an anthropic.Anthropic client pointed at the tare proxy local proxy
openai_with_tare(client_kwargs=None, base_url="http://127.0.0.1:8787") Returns an openai.OpenAI client pointed at the tare proxy local proxy

The first six compress in-process via the Rust binding — no proxy process required. anthropic_with_tare and openai_with_tare route through the local proxy instead; start it first with tare-proxy.

JS / TS — tare-ai

const { withTare, tareMiddleware, startProxy, tareBaseUrl } = require("tare-ai");
// or ESM / TypeScript:
import { withTare, tareMiddleware, startProxy, tareBaseUrl } from "tare-ai";

All JS / TS adapters route through the local proxy (tare-proxy, default port 8787).

Export What it does
withTare(clientOptions?, port?) Merges the proxy baseURL into SDK client options — works with both the Anthropic and OpenAI JS SDKs: new Anthropic(withTare({ apiKey: "..." }))
tareMiddleware(port?) Returns a LanguageModelV1Middleware-shaped object for the Vercel AI SDK; attaches the proxy URL as _tare_base. Actual HTTP routing still requires withTare on the underlying provider client
startProxy(opts?) Spawns the vendored tare-proxy binary and returns { child, baseUrl, stop }
tareBaseUrl(port?) Returns "http://127.0.0.1:<port>" (utility)