Hebline

io.github.HeblinePublished by GitHub account Heblinev0.2.1Updated Nov 10, 2026

Intelligent API routing for AI agents. Best service, best price, free alternatives.

Overview

AI-generated overview

Routes an assistant's API calls (LLMs, geocoding, translation, search, weather and more) to a free or paid provider and returns the result with cost and latency

What you can do
Hebline is a local MCP server that acts as a broker for external APIs. Its execute tool picks a service for a requested capability, makes the call and returns the result plus metadata such as service, cost and latency; compare shows all services for a capability with their scores, and categories lists supported capabilities. It covers nine categories and twenty services, including LLMs, geocoding, translation, web scraping, currency, OCR, weather, web search and news, preferring free providers unless a paid key is configured.
When to use it
Use it when an assistant needs several kinds of external lookups and you want one interface that prefers free providers and reports what each call cost. It suits agents that would otherwise call many separate APIs and where you want visibility into spend.
Requirements
Runs as a local stdio process, installed via npm (npx or global install) and configured in the client's MCP config. Node.js is needed. Free providers work without keys; paid providers are enabled by setting DEEPL_API_KEY, GOOGLE_MAPS_API_KEY or LIBRETRANSLATE_API_KEY, and the README also lists other provider keys. Network access is required. Without a local LLM key, LLM calls go through the Hebline proxy with a daily free limit.
Before you install
Paid providers are billed to your own accounts when their keys are set, and the README notes LLM calls may route through the Hebline proxy when no local key is set. Calls are logged to a local file with metadata such as capability, service, latency and cost, though the README says no call content is stored. Requests and inputs are sent to third-party providers, so avoid sending sensitive data.

Installation

In SourceWeft

  1. Open Hebline in the dashboard and add it to a workspace.
  2. Enable the server for the chats that should use its tools.

Desktop only via STDIO. STDIO servers start a local process, so they need the SourceWeft desktop host.

Other MCP clients

Follow the launch instructions in the repository.

README

Hebline MCP Server

Your agents overpay for every API call. We fix that.

Hebline routes every API call — including LLM calls — to the best service at the right price. Free when it's enough. Paid when it matters. It knows the difference.

Every other router earns a margin on your paid calls. Routing you to free alternatives kills their revenue. No margin on your API calls. Ever.

Why Hebline?

Your agents are bleeding money. One task triggers 5–10 paid API calls across different providers. No transparency, no cost control. Hebline fixes that:

  • Route free first — Most calls don't need the best model. Hebline learns precisely when it matters — and keeps learning as the market changes.
  • No margin. Honest routing. — We don't earn when you pay more. So we're the only router built to actually save you money.
  • Provider Abstraction — Your agent says what it needs ("geocode this address"), not which service to use. Swap providers without changing agent code.
  • Cost Transparency — Every call is logged with service used, latency, and cost. Know exactly what your agents spend.
  • Learns from usage — Hebbian learning strengthens what works, weakens what doesn't. Your broker gets smarter every day.
  • BYOK (Bring Your Own Key) — Paid services use your API keys via environment variables. No key? The service is automatically excluded from routing.
  • GDPR compliant — Only anonymized metadata logged. No API call content stored. Self-hosted option for zero data leaving your network.
  • Open Source — Core MCP server is MIT licensed. Community-driven adapter system.

How It Works

Your AI Agent ←→ Hebline MCP Server ←→ Best API (Nominatim, DeepL, Google Maps, ...)                        │                   Smart Routing                   Cost Logging                   Provider Scoring

Your agent connects to Hebline as an MCP server. Instead of calling APIs directly, it uses Hebline's tools — execute, compare, or categories. Hebline scores all available services, picks the best one, makes the call, and returns the result with full metadata.

Available MCP Tools

ToolDescription
executeRoute to the best service and make the API call. Returns result + metadata (service, cost, latency).
compareShow all available services for a capability with scores. See what's available before committing.
categoriesList all supported capabilities and their services.

Quick Start

Add to Claude Desktop

Add to claude_desktop_config.json:

json
{  "mcpServers": {    "hebline": {      "command": "npx",      "args": ["-y", "-p", "@hebline.ai/mcp-server", "hebline-mcp"]    }  }}

Add to Claude Code

Add to .mcp.json:

json
{  "mcpServers": {    "hebline": {      "command": "hebline-mcp"    }  }}

Add to Cursor

Add to .cursor/mcp.json:

json
{  "mcpServers": {    "hebline": {      "command": "npx",      "args": ["-y", "-p", "@hebline.ai/mcp-server", "hebline-mcp"]    }  }}

Add to Windsurf

Add to ~/.codeium/windsurf/mcp_config.json:

json
{  "mcpServers": {    "hebline": {      "command": "npx",      "args": ["-y", "-p", "@hebline.ai/mcp-server", "hebline-mcp"]    }  }}

Add to VS Code (Copilot)

Add to .vscode/mcp.json:

json
{  "servers": {    "hebline": {      "type": "stdio",      "command": "npx",      "args": ["-y", "-p", "@hebline.ai/mcp-server", "hebline-mcp"]    }  }}

Install globally

bash
npm install -g @hebline.ai/mcp-server

With paid services (optional)

Set environment variables for any paid providers you want to use:

env
GOOGLE_MAPS_API_KEY=your-key-hereDEEPL_API_KEY=your-key-hereLIBRETRANSLATE_API_KEY=your-key-here

No keys? No problem — Hebline routes to free alternatives automatically.

Supported Services

CategoryFreePaid (BYOK)
LLMsGroq (Llama 3.3 70B), Google Gemini FlashOpenAI GPT-4o-mini (OPENAI_API_KEY)
GeocodingNominatim (OpenStreetMap)Google Maps (GOOGLE_MAPS_API_KEY)
TranslationMyMemoryDeepL (DEEPL_API_KEY), LibreTranslate (LIBRETRANSLATE_API_KEY)
Web ScrapingFetch ScraperFirecrawl (FIRECRAWL_API_KEY)
CurrencyExchangeRate-APIFixer.io (FIXER_API_KEY)
OCROCR.spaceGoogle Vision (GOOGLE_VISION_API_KEY)
WeatherOpen-MeteoOpenWeatherMap (OPENWEATHERMAP_API_KEY)
Web SearchDuckDuckGoBrave Search (BRAVE_API_KEY)
NewsHackerNewsNewsAPI.org (NEWSAPI_KEY)

9 categories, 20 services. Free services work instantly — no API key needed. LLMs route through the Hebline proxy when no local key is set (50 free calls/day).

Example

An agent asks: "Geocode the Brandenburg Gate in Berlin"

Hebline receives:

json
{  "capability": "geocoding",  "input": { "query": "Brandenburger Tor, Berlin" },  "constraint": "free"}

Hebline responds:

json
{  "success": true,  "data": {    "lat": 52.5163,    "lon": 13.3777,    "displayName": "Brandenburger Tor, Pariser Platz, Berlin, 10117, Deutschland"  },  "meta": {    "service": "Nominatim (OpenStreetMap)",    "costUsd": 0,    "latencyMs": 258,    "score": 0.702,    "free": true  }}

The agent got coordinates, knows it was free, and Hebline logged the call for future analysis.

Architecture

mcp-server/├── src/│   ├── index.ts              # MCP server entry point (stdio transport)│   ├── types.ts              # Shared TypeScript types│   ├── registry.ts           # Service definitions (capabilities, costs, scores)│   ├── router.ts             # Weighted scoring engine (Hopfield-ready)│   ├── logger.ts             # Append-only JSONL call log (~/.hebline/calls.jsonl)│   ├── adapters/             # One adapter per service│   │   ├── nominatim.ts      # Free geocoding│   │   ├── google-maps.ts    # Paid geocoding (BYOK)│   │   ├── mymemory.ts       # Free translation│   │   ├── libretranslate.ts # Paid translation (BYOK)│   │   └── deepl.ts          # Paid translation (BYOK)│   └── tools/                # MCP tool definitions│       ├── execute.ts        # Route + call best service│       ├── compare.ts        # Score all services│       └── categories.ts     # List capabilities

Call Logging

Every API call is logged to ~/.hebline/calls.jsonl:

json
{"timestamp":"2026-03-29T09:36:37Z","capability":"geocoding","serviceId":"nominatim","latencyMs":212,"success":true,"costUsd":0}

No content is logged — only metadata. This data will power Hebbian Learning in future versions.

Roadmap

  • Core MCP server with stdio transport
  • Weighted scoring router
  • Geocoding adapters (Nominatim, Google Maps)
  • Translation adapters (MyMemory, LibreTranslate, DeepL)
  • BYOK key management
  • Append-only call logging
  • CI/CD with GitHub Actions
  • Hebbian Learning — router learns from call history
  • Hopfield network scoring (replaces weighted scoring)
  • More categories (web scraping, currency, OCR, email)
  • Community adapter system
  • SSE transport for remote deployments
  • Web dashboard for cost analytics
  • Budget alerts and spending limits
  • Multi-agent cost attribution

Contributing

Contributions are welcome! Adding a new adapter is straightforward — implement the ServiceAdapter interface and register it.

bash
git clone https://github.com/hebline/mcp-server.gitcd mcp-servernpm installnpm run buildnpm test

License

MIT


Built by Hebline — Route free first. Only pay when necessary.

Source: README.md at commit 548ed23

Tools

0
SourceWeft has not started this package in its sandbox yet, so its tools are not known.

Version history

1
  1. v0.2.1LatestSep 16, 2026

Score

50% of the applicable checks met

How complete this listing is, computed from the registry entry, the endpoint and public sources. Not a rating of the server and not a statement that it is safe or working today: reachability is a dated check, and no installs or reviews go into the score.

  • Reached in the last 30 days25 pts

    Keep the endpoint answering the MCP handshake; SourceWeft probes it every 30 days.

  • Auth declared10 pts

    Declare the credential headers in the registry entry, or keep the endpoint answering the probe so its auth can be classified.

  • Tools listed15 pts

    Answer tools/list without a credential, or ship a package that starts in a sandbox without one.

  • DescriptionRequired15 pts
  • Logo10 pts
  • Documentation15 pts
  • License10 pts