NEW MCP SERVER · v0.1

Hand the catalog
to your agent.

@abliteration/mcp is a Model Context Protocol server that lets Claude Code, Cursor, Windsurf and Claude Desktop talk to the abliteration.org catalog directly - eighteen read-only tools, metered in the same credits as the public API, no extra billing surface.

Live recording. The agent decides which tools to call and when - every call is metered the same as a direct HTTP request.

three moving parts

How it works.

01
Your editor spawns a local process

When you open Claude Code or Cursor, it runs npx -y @abliteration/mcp on your machine. The package is a ~500 KB npm download that caches locally after the first install. No server is dialed yet; the process is a local bridge, nothing more.

02
The agent discovers 18 tools

The bridge announces its toolkit over the MCP standard (stdio). From here on, your agent reasons about them the same way it reasons about reading a file: it picks the tool that fits the question and fills the arguments.

03
Each call hits the public API

A tool call becomes an HTTPS request to abliteration.org/api/v1/* with your key in the Authorization header. Credits are debited server-side, the response flows back through the bridge to the agent, the agent explains it to you. The key never leaves your machine except as a request header.

No new service to host, no SaaS between you and the catalog, no second wallet. The server you already pay is the one your agent is talking to.

four hosts, one package

Install.

Pick your editor. Each snippet uses the same npm package; only the configuration surface differs. Replace abl_live_... with your real key from your account page.

Claude Code

command line, auto-written
claude mcp add abliteration \
  --env ABLITERATION_API_KEY=abl_live_... \
  -- npx -y @abliteration/mcp

Writes the entry into Claude Code's own config. Restart the session afterwards.

Cursor

~/.cursor/mcp.json
{
  "mcpServers": {
    "abliteration": {
      "command": "npx",
      "args": ["-y", "@abliteration/mcp"],
      "env": {
        "ABLITERATION_API_KEY": "abl_live_..."
      }
    }
  }
}

Create or edit the file, drop in the snippet, restart Cursor.

Windsurf

~/.codeium/windsurf/mcp_config.json
{
  "mcpServers": {
    "abliteration": {
      "command": "npx",
      "args": ["-y", "@abliteration/mcp"],
      "env": {
        "ABLITERATION_API_KEY": "abl_live_..."
      }
    }
  }
}

Same JSON. Windsurf picks it up on next session.

Claude Desktop

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
{
  "mcpServers": {
    "abliteration": {
      "command": "npx",
      "args": ["-y", "@abliteration/mcp"],
      "env": {
        "ABLITERATION_API_KEY": "abl_live_..."
      }
    }
  }
}

Add the abliteration entry, quit and reopen the app.

eighteen tools

What your agent gets.

Every tool is read-only. Every tool prints its credit cost in its own description, so an agent that reasons about budget can plan ahead instead of discovering the bill mid-session.

Catalog 11

The parts anyone could in principle build on top of Hugging Face - but faster, and with the taxonomy already filled in.

  • search_models 5-300+ cr

    Filter the catalog by method, family, author, downloads, license, dates. Sort by any field.

  • get_model 10 cr

    One model with files, benchmarks and an author summary. Delisted rows still return - flagged.

  • list_methods 5-15 cr

    The M-code taxonomy with live model counts.

  • get_method 1 cr

    One method: definition, count, top authors, sample models.

  • list_authors 5-15 cr

    Authors ranked by downloads or model count.

  • get_author 1 cr

    One author with aggregate totals and their top models.

  • search_datasets 5-15 cr

    Alignment datasets filterable by category, author and license.

  • get_dataset 1 cr

    One dataset with files, and its passport when it has one.

  • list_indices 20 cr

    All four flagship indices with recent DAI history.

  • get_index 10 cr

    One index with its full history series.

  • get_timeline 50 cr

    Curated field events: launches, takedowns, policy jolts.

Change streams 7 · the moat

The feeds Hugging Face does not give you - deletions, README drift, download anomalies. The reason the API exists.

  • list_delisted 5 cr

    Models pulled from Hugging Face since your last check. The feed Hugging Face does not give you.

  • get_deletions_summary 20 cr

    24-month deletion analysis rolled up. One call replaces half a dozen queries.

  • list_deletions 5 cr

    Paged backlog of the deletion archive. Filter by month, peak downloads.

  • list_readme_drifts 5 cr

    Every commit that changed a model's README, with size_delta per commit.

  • get_readme_at 10 cr

    Wayback: the full README markdown at any commit sha.

  • get_download_anomalies 15 cr

    Models whose downloads just spiked 2× or dropped by half. Scored by absolute delta.

  • get_download_history 5 cr

    Per-model download timeline - the raw signal under the anomaly feed.

one wallet, no second bill

Pricing and the key.

Same credits as the API

Each tool call is one HTTP request to /api/v1/*. The server bills it under the same rate card as a curl would - the credit cost on this page is the cost at your wallet. No MCP tax, no overhead, no proxy fee. Credit balance is echoed back to the agent so it can report it on request.

Pricing table: terms, section 7.

Where the key sits

The key lives in your editor's config, as an environment variable passed to the local process. It never reaches abliteration.org except in the Authorization: Bearer header of each request. Revoke a leaked key the same way you revoke any API key - on your account page.

Need a key? Create one. New accounts get free starter credits.

open source

The code.

TypeScript, Node 18+, strict mode, no any. Depends on @modelcontextprotocol/sdk and zod, nothing else. Published under @abliteration/mcp on npm; source lives in its own public repo at github.com/neymark/abliteration-mcp.

Open in Abliteration