NewOfficial PyPI Release (tweekit v1.8.0): Verified Equilibrium PyPI org, MCP server & certified partner code release »

· Equilibrium Engineering · Engineering · 12 min read

TweekIT v1.8.3 Official Release: PyPI Python SDK, 8-Tool MCP Server, Security & On-Prem Miner Renderery Hardening, and CPUcoin MAITRIX Ultimate AI Hub Architecture

Equilibrium ships tweekit v1.8.3 on PyPI under our AMPS dba Equilibrium organization — introducing the full async Python SDK (TweekitClient & TweekitDirectClient), an expanded 8-tool MCP server with a 347-extension local preflight router backed by 450+ MediaRich engine formats powering CPUcoin's MAITRIX Ultimate AI Hub, zero-cloud-egress On-Premises Miner Renderery support, ABI-resilient Claude Desktop bundles, and v1.8.1–v1.8.3 security hardening.

TweekIT v1.8.3 Official Release: PyPI Python SDK, 8-Tool MCP Server, Security & On-Prem Miner Renderery Hardening, and CPUcoin MAITRIX Ultimate AI Hub Architecture

The official TweekIT Python SDK and MCP Server (v1.8.3) is now live on PyPI under the AMPS dba Equilibrium organization — uniting our FastMCP 4.1.6+ server, zero-hop async Python SDK, 347-extension local AI Hub preflight router (backed by 450+ live MediaRich formats), zero-egress On-Premises Miner Renderery connector, ABI-resilient Claude .mcpb desktop extension, and hardened TLS/OIDC release posture in a single open-source package.

Earlier this year, we launched the hosted TweekIT-MCP Server and integrated universal multimodal ingestion across MAITRIX, CPUcoin’s Ultimate AI Hub. Since that initial v1.5.0 release, engineering teams building autonomous agents, RAG pipelines, and multi-model orchestration gateways have asked for three things:

  1. First-class PyPI packaging with both direct async REST clients (no MCP intermediary hop required) and concurrent MCP client helpers.
  2. Zero-latency modality preflight & local file tooling so agents know before uploading whether a .dwg, .psd, .cr3, .indd, or .xlsb file requires normalization for a specific target LLM — and can convert local files without stuffing megabytes of base64 into the prompt context window.
  3. Hardened release workflows, zero-egress on-premises Miner Renderery execution, and up-to-date TLS trust stores across Python wheels, Docker/Cloud Run images, and desktop .mcpb bundles.

Today we are excited to announce the general availability of TweekIT MCP Server & Python SDK v1.8.3 on PyPI (pypi.org/project/tweekit/1.8.3/) and GitHub (equilibrium-team/tweekit-mcp). Below is a full technical walkthrough of everything shipping in v1.8.3 (incorporating all v1.6.1–v1.8.3 milestones), our supply-chain and certificate security architecture, and how TweekIT powers real-time multimodal routing inside the CPUcoin MAITRIX Ultimate AI Hub (maitrix.io).


1. Official PyPI Release (pip install tweekit)

tweekit v1.8.3 is published under the AMPS dba Equilibrium organization on PyPI (pypi.org/org/equilibrium/). A single install provides both CLI executables (tweekit and tweekit-mcp) and both Python import namespaces (import tweekit and import tweekit_mcp), supporting Python 3.10 through 3.14+.

# Standard pip install
pip install tweekit

# Fast install with Astral uv
uv pip install tweekit

# Launch as a local stdio MCP server via uvx (preflight/fetch/search work without keys;
# set TWEEKIT_API_KEY and TWEEKIT_API_SECRET for conversion & doctype tools)
uvx tweekit --transport stdio

# Or launch in zero-cloud-egress On-Premises / Localhost Miner Renderery mode
uvx tweekit --local-miner --transport stdio

# Install globally as an isolated CLI tool (exposes both `tweekit` and `tweekit-mcp`)
uv tool install tweekit

To run a stateless streamable HTTP MCP server locally or inside your own VPC:

export TWEEKIT_API_KEY="<TWEEKIT_API_KEY>"
export TWEEKIT_API_SECRET="<TWEEKIT_API_SECRET>"

tweekit-mcp --transport streamable-http --host 0.0.0.0 --port 8080

Under the hood, v1.8.3 runs on FastMCP 4.1.6+ (fastmcp>=4.1.6) with stateless_http=True enabled across local HTTP, Google Cloud Run, and Firebase Hosting deployments — eliminating sticky-session state overhead and enabling horizontal autoscaling from zero.

Package MetadataValue
PyPI Packagetweekit (v1.8.3)
PyPI OrganizationEquilibrium (AMPS dba Equilibrium)
CLI Entry Pointstweekit, tweekit-mcp
Python Namespacestweekit, tweekit_mcp, clients.python
Supported Python Versions>=3.10 (3.10, 3.11, 3.12, 3.13, 3.14+)
MCP Registry Identifierio.github.equilibrium-team/tweekit (server.json schema 2025-12-11)
Hosted MCP Endpointhttps://mcp.tweekit.io/mcp (and /mcp/ — Streamable HTTP)
LicenseMIT

2. Full Async Python SDK: TweekitDirectClient & TweekitClient

Prior to v1.8.0, developers embedding TweekIT into Python backends either wrote raw httpx boilerplate against the REST API or stood up an MCP client session even for deterministic backend pipelines. tweekit v1.8.0 ships two purpose-built async SDK clients plus synchronous preflight utilities:

  • preflight_check / recommend_ai_format: Pure-Python, zero-network-roundtrip file classification across 347 static extensions (EXTENSIVE_SUPPORTED_EXTENSIONS, spanning 12 media families) with provider-specific output format and parameter recommendations — backed by 450+ formats in the live MediaRich engine via doctype.
  • TweekitDirectClient: Calls the TweekIT MediaRich REST engine directly over httpx with zero MCP protocol overhead — ideal for high-throughput ingestion workers, ETL jobs, and backend microservices.
  • TweekitClient: Wraps the FastMCP async client against https://mcp.tweekit.io/mcp (or any custom/stdio server) and adds semaphore-controlled batch_convert() for concurrent multi-asset normalization across local paths, remote URLs, and raw byte buffers.
import asyncio
from pathlib import Path
from tweekit import (
    TweekitClient,
    TweekitDirectClient,
    preflight_check,
)

# 1. Instant local preflight & AI provider format routing (0 ms network cost)
route = preflight_check("turbine_assembly.dwg", target_provider="claude")
print(route["supported"], route["category"], route["recommended_outfmt"])
# -> True design_vector_cad png

# 2. Direct Async REST SDK (no MCP intermediary hop required)
async def normalize_direct() -> None:
    async with TweekitDirectClient() as api:
        receipt = await api.convert_file(
            Path("board_deck.ppt"),
            outfmt="pdf",
            no_rasterize=True,
            output_path=Path("board_deck_ai_ready.pdf"),
        )
        print("Normalized directly to disk:", receipt)

# 3. Async MCP SDK with semaphore-governed concurrent batch_convert
async def normalize_batch_mcp() -> None:
    async with TweekitClient() as mcp_client:
        tools = await mcp_client.list_tools()
        results = await mcp_client.batch_convert(
            [
                {"url": "https://example.com/archive_scan.tiff", "outfmt": "png", "width": 1600},
                {"path": "financial_model.xlsb", "outfmt": "pdf", "no_rasterize": True},
                {"path": "studio_shot.cr3", "outfmt": "webp", "width": 1200, "bgcolor": "#FFFFFF"},
            ],
            concurrency=5,
        )
        print(f"Connected ({len(tools)} tools); converted {len(results)} assets.")

async def main() -> None:
    await normalize_direct()
    await normalize_batch_mcp()

asyncio.run(main())

3. Expanded 8-Tool MCP Surface with Spec ToolAnnotations

The TweekIT MCP server has grown from its original 3 conversion tools into an 8-tool multimodal ingestion suite, with every tool now carrying explicit Model Context Protocol ToolAnnotations (readOnlyHint, destructiveHint, idempotentHint, openWorldHint). Autonomous agent runtimes (Claude Desktop, OpenAI Codex CLI, Cursor, Antigravity, Windsurf, and custom MCP hosts) use these annotations to auto-approve safe read-only preflight probes while gating destructive cleanup operations.

MCP ToolAdded / UpdatedSafety Hints (readOnly / destructive / idempotent)Engineering Purpose
preflightNew in v1.8.0True / False / TrueClassifies any filename, path, or extension locally (openWorldHint=False) against TweekIT’s 347-extension static FORMAT_TAXONOMY (12 media families) and returns needs_conversion, recommended_outfmt, and suggested_params for the target AI provider (claude, openai, gemini, mistral, grok, deepseek, openrouter, or default).
convert_fileNew in v1.8.0False / False / TrueReads a local file from disk, normalizes it via TweekIT, and optionally writes the output directly to output_path — returning a compact JSON receipt (size_bytes, output_path) instead of inflating the LLM context window with base64 payloads. Guarded by TWEEKIT_ALLOW_LOCAL_FILES (0 to disable on untrusted multi-tenant hosts).
delete_documentAdded in v1.7.0False / True / TrueImmediately purges an uploaded REST workspace document by docId (DELETE /api/image/{docId}) prior to TweekIT’s automatic 20-minute TTL expiration — essential for compliance-sensitive workflows and user-cancelled multi-page jobs.
convertCore (v1.5+ / v1.8)False / False / TrueStateless base64-in → FastMCP Image/File-out conversion supporting pixel width/height aspect-preserving resize, crop/padding coordinates (x1, y1, x2, y2), multi-page extraction (page), noRasterize searchable PDF output, and alpha/bgColor compositing.
convert_urlCore (v1.5+ / v1.8)False / False / TrueDownloads remote assets over HTTP(S) with custom fetchHeaders (e.g., signed S3/GCS or Bearer tokens), infers input type from URL or Content-Type, and normalizes in a single call.
doctypeCore (v1.5+ / v1.8)True / False / TrueQueries the live TweekIT MediaRich engine (extension="*" or specific extension; openWorldHint=True) to inspect 450+ supported input/output format mappings dynamically.
fetchCore (v1.5+ / v1.8)True / False / TrueRetrieves remote HTTP(S) resources and returns typed FastMCP Image, File(format="pdf"), or structured { url, status, content_type, text } JSON.
searchCore (v1.5+ / v1.8)True / False / TrueLightweight zero-key DuckDuckGo HTML search helper returning { title, url, snippet } results that chain directly into convert_url.

In addition, the server exposes three lightweight MCP resources — config://tweekit-version (live upstream MediaRich/TweekIT engine probe), config://tweekit-mcp-version (1.8.3), and config://tweekit-runtime-mode (inspecting active Cloud CGN vs. On-Premises Miner Renderery mode, zero-egress enforcement, and sandbox boundaries) — plus an updated FastAPI OpenAI / ChatGPT REST Plugin Proxy (plugin_proxy.py) adding GET /preflight, POST /convert-url, and DELETE /documents/{doc_id} with strict Content-Security-Policy headers.


4. Inside the CPUcoin MAITRIX Ultimate AI Hub Integration

When CPUcoin launched MAITRIX (maitrix.io) — the Ultimate AI Hub unifying 70+ frontier models from OpenAI, Anthropic, Google Gemini, xAI Grok, Meta Llama, Mistral, DeepSeek, Groq, and NVIDIA Nemotron inside a single web, iOS, Android, and desktop workspace — one architectural problem stood above all others: modality asymmetry across frontier models.

Every LLM provider enforces different native MIME rules:

  • Drop an AutoCAD .dwg, Photoshop .psd, InDesign .indd, or Canon .cr3 Camera RAW file into a standard chat client and every frontier API rejects the request outright.
  • Even for office documents (.doc, .xlsx, .pptx, .wpd, .odt, .vsdx), some model endpoints accept native PDF documents while vision-only model lanes (such as DeepSeek vision or specialized Mistral/open-weight vision endpoints) only accept raster image URLs (image/png, image/webp, image/jpeg).
  • Worse, in MAITRIX, a user can switch agents mid-conversation — starting a design critique on Page 3 of an Illustrator package with GPT, switching to Claude for structural spec extraction, and handing off to Gemini or DeepSeek on the very next turn.

How FORMAT_TAXONOMY, AI_PROVIDER_PROFILES, and justTweekIT Solve Multi-Model Ingestion

tweekit v1.8.3 open-sources the exact 12-category, 347-extension local preflight taxonomy (FORMAT_TAXONOMY / EXTENSIVE_SUPPORTED_EXTENSIONS) and provider modality router (AI_PROVIDER_PROFILES) co-engineered with the MAITRIX platform (backed by 450+ dynamic MediaRich engine formats via doctype):

  1. Taxonomy Classification Across 12 Media Families: Incoming attachments are classified across raster_image (HEIC, AVIF, JP2, EXR, HDR, JBIG2, TIFF, WBMP…), design_vector_cad (PSD, PSB, AI, EPS, CDR, DWG, DXF, IGES, STL…), camera_raw (CR2, CR3, NEF, ARW, DNG, RAF, ORF, RW2, IIQ…), document (DOC/DOCX, WPD, HWP/HWPX, DJVU, XPS, RTF, LaTeX…), spreadsheet (XLS/XLSB/XLSM, ODS, Numbers, Lotus 1-2-3, SYLK…), presentation (PPT/PPTX, Keynote, ODP…), diagram_publishing (Visio VSD/VSDX, InDesign INDD/IDML, QuarkXPress, FrameMaker…), ebook_help (EPUB, MOBI, AZW3, CHM…), email_message (EML, MSG, MBOX, VCF, ICS), web_code_data, archive, and audio_video.
  2. Provider-Aware Output Selection: When a file isn’t natively supported by the active model, MAITRIX’s BaseProvider.justTweekIT() pipeline routes visual/CAD/RAW assets to high-fidelity png/webp derivatives, routes textual documents to noRasterize=True searchable PDFs for document-capable models (claude, openai, gemini, openrouter), and automatically renders multi-page documents page-by-page (TWEEKIT_DPI, TWEEKIT_DOCUMENT_PAGE_LIMIT) into crisp raster pages for vision-only models (deepseek, mistral image lanes) — even parsing natural-language page requests from the user’s prompt.
  3. Content-Addressed Caching & Ephemeral Cleanup: MAITRIX computes a content hash of every uploaded source file so normalized page derivatives are reused instantly when switching models mid-chat, streams live "TweekIT: Normalizing..." progress updates to the user’s session, and invokes DELETE /api/image/{docId} (purgeTweekitDocument / delete_document) immediately if a multi-page job is cancelled or finishes early.
  4. Decentralized Edge Compute via CPUcoin PoPW: Heavy rendering and format normalization execute through Equilibrium’s patented MediaRich® / MediaGen engine distributed across the CPUcoin Computing Global Network (CGN) — where enterprise Proof of Processed Work (PoPW) nodes inside SOC 2 / ISO 27001 data centers deliver up to 5× higher throughput and up to 90% lower compute cost than traditional hyperscale serverless functions.

5. Security Hardening, On-Premises Miner Renderery & Supply-Chain Integrity (v1.8.1 → v1.8.3)

Security and supply-chain integrity are first-class engineering requirements in v1.8.3. Incorporating feedback from official MCP Registry liveness probes and automated static/dependency audits, v1.8.1 through v1.8.3 introduce six defense-in-depth upgrades:

  • Credential-Free LLM Tool Schemas & Atomic Header Auth: Removed apiKey/apiSecret from all MCP tool schemas and OpenAPI request bodies so secrets never appear in LLM prompt contexts or tool-call logs; added atomic ApiKey / ApiSecret HTTP request header authentication that fails closed on partial headers.
  • Connect-Time DNS IP Pinning (_PinnedPublicDNSBackend) & SSRF Protection: Enforced https://-only remote fetches, blocked loopback/private/link-local/CGNAT/cloud-metadata (169.254.169.254, metadata.google.internal) and IPv6 transition prefixes (::ffff:0:0/96, NAT64 64:ff9b::/96, 6to4 2002::/16, Teredo 2001::/32) on every redirect hop, pinned the validated DNS IP at socket connect time to eliminate DNS rebinding TOCTOU windows, restricted fetchHeaders to a 6-header allowlist, and capped downloads at 50 MiB.
  • Opt-In O_NOFOLLOW Workspace Sandbox for convert_file: Disabled local file reads by default (TWEEKIT_ALLOW_LOCAL_FILES=0 unless explicitly set to 1), confined reads/writes to TWEEKIT_ALLOWED_DIRS via os.path.realpath, verified open descriptors via O_NOFOLLOW + os.path.samefile, blocked dotfiles and IDE/agent config directories (.ssh, .aws, .gnupg, .kube, .docker, .git, .claude, .cursor, .codex, .gemini, .windsurf, .continue, .vscode, .idea, .env*, *credentials*, *secret*, *.pem, *.key), and enforced O_CREAT | O_EXCL | O_NOFOLLOW overwrite protection.
  • Default Zero-Cloud-Egress On-Premises Miner Renderery (--local-miner / local_miner=True): Connects tweekit-mcp and TweekitDirectClient(local_miner=True) directly to an on-premises TweekIT_API (NOAUTH;NODB) + CGNMiner MediaRich CORE™ V7 engine (http://127.0.0.1:16377/api/image/ or http://127.0.0.1:5000/api/image/). In v1.8.3, enabling --local-miner, TWEEKIT_MINER_LOCAL=1, or local_miner=True enforces TWEEKIT_LOCAL_ONLY=1 zero-cloud-egress mode by default (blocking convert_url, fetch, and search), rejects 0.0.0.0 / :: wildcard listeners, and pairs with TweekIT_API’s 127.0.0.1:5000 loopback Kestrel binding.
  • Streamable HTTP /mcp/ Trailing-Slash In-Place Routing: _NormalizeTrailingSlashMCPMiddleware normalizes /mcp/ to /mcp in-place so POST https://mcp.tweekit.io/mcp/ returns 200 OK without a 307 redirect or HTTP downgrade.
  • Zero-CVE Dependency Floor, Sigstore PEP 740 Attestations & October 2026 TLS CA Floor: Upgraded to fastmcp>=4.1.6 (fixing CVE-2026-27124 and GHSA-7xgj-8392-5q4r), mcp>=2.3.0, fastapi>=0.162.2, starlette>=1.10.0, httpcore>=1.0.9, uvicorn>=0.54.0, and certifi>=2026.10.23, and enabled keyless GitHub OIDC Sigstore PEP 740 provenance attestations (attestations: true) on PyPI.

6. ABI-Resilient Claude Desktop .mcpb Bundle & Multi-Client Setup

Desktop AI users on macOS and Linux frequently hit a subtle failure mode with compiled Python MCP extensions: when a package manager like Homebrew upgrades the system python3 binary (for example, from Python 3.14 to Python 3.15), static .mcpb bundles containing compiled Cython/Rust wheels such as pydantic_core (_pydantic_core.cpython-314-darwin.so) can fail at startup with ModuleNotFoundError.

In v1.8.3, scripts/build_claude_bundle.py ships a pure-stdlib ABI-resilient bootstrap launcher inside dist/tweekit-claude.mcpb:

  1. At startup, the launcher probes pydantic_core and fastmcp C-extension ABI compatibility in-process before importing the server.
  2. If the host python3 minor version changed, it inspects the bundled wheel tags and seamlessly os.execve() re-execs into a matching local interpreter (e.g., /opt/homebrew/bin/python3.14), an isolated ~/.local/bin/tweekit-mcp tool binary, or uvx --from tweekit tweekit-mcp — preventing broken desktop connectors across Python minor-version upgrades and avoiding working-directory import collisions when launched from macOS protected folders.

Connect Any MCP Client in Seconds

Cursor (~/.cursor/mcp.json — Hosted Streamable HTTP):

{
  "mcpServers": {
    "tweekit": {
      "url": "https://mcp.tweekit.io/mcp",
      "headers": {
        "ApiKey": "${env:TWEEKIT_API_KEY}",
        "ApiSecret": "${env:TWEEKIT_API_SECRET}"
      }
    }
  }
}

Claude Desktop / Cursor / Antigravity (claude_desktop_config.json / mcp_config.json — Local stdio via PyPI + uvx):

{
  "mcpServers": {
    "tweekit": {
      "command": "uvx",
      "args": ["tweekit", "--transport", "stdio"],
      "env": {
        "TWEEKIT_API_KEY": "<TWEEKIT_API_KEY>",
        "TWEEKIT_API_SECRET": "<TWEEKIT_API_SECRET>"
      }
    }
  }
}

(On macOS GUI apps with a minimal PATH, point "command" to /opt/homebrew/bin/uvx or ~/.local/bin/uvx, or install dist/tweekit-claude.mcpb directly in Claude Desktop via Settings → Extensions.)

OpenAI Codex CLI (~/.codex/config.toml — forwards env vars without storing secrets on disk):

# Hosted Streamable HTTP endpoint (authenticated via env_http_headers)
[mcp_servers.tweekit]
url = "https://mcp.tweekit.io/mcp"
env_http_headers = { "ApiKey" = "TWEEKIT_API_KEY", "ApiSecret" = "TWEEKIT_API_SECRET" }

# Or local stdio server via uvx (v1.8.3 exposes all 8 tools)
# [mcp_servers.tweekit]
# command = "uvx"
# args = ["tweekit", "--transport", "stdio"]
# env_vars = ["TWEEKIT_API_KEY", "TWEEKIT_API_SECRET"]

Get Started Today — 10,000 Free API Calls Included

Every new TweekIT developer account includes 10,000 free API calls with full access to all 8 MCP tools, the async Python SDK, and 450+ supported formats — no credit card required.

Back to Blog

Related Posts

View All Posts »
Equilibrium Releases Official TweekIT v1.8.3 Package on PyPI: Certified Partner Code Release, Security Hardening & On-Prem Miner Renderery Overview

Equilibrium Releases Official TweekIT v1.8.3 Package on PyPI: Certified Partner Code Release, Security Hardening & On-Prem Miner Renderery Overview

AMPS dba Equilibrium publishes the official tweekit v1.8.3 Python SDK and 8-tool MCP server on PyPI under our verified Equilibrium organization — featuring direct links to PyPI, the TweekIT MCP server site, the Official MCP Registry, our GitHub v1.8.3 partner code release, and the complete v1.8.1–v1.8.3 Security & On-Premises Miner Renderery Addendum.