Kryfto – Self-hosted MCP server with 42 tools for AI agents

TL;DR

TL;DR: Kryfto is an open-source headless browser runtime that exposes 42 MCP tools for AI agents, letting Claude Code, Cursor, and Codex browse, search, extract, and interact with the live web under a self-hosted Docker setup.

Source and Accuracy Notes

⚠️ This section is MANDATORY. All links must be verified from actual source, not guessed.

• Project page: github.com/ExceptionRegret/Kryfto
• Source repository: github.com/ExceptionRegret/Kryfto — README.md read in full (v3.8.0)
• License: Apache 2.0 — verified
• HN launch thread: news.ycombinator.com/item?id=47252539
• Current version: v3.8.0 (verified via GitHub releases page)

What Is Kryfto?

Kryfto is a self-hosted web browsing backend for AI agents. It solves the fundamental gap between LLMs and real web interaction: models trained on static data cannot browse live sites, fill forms, solve CAPTCHAs, or extract structured content from dynamic pages.

Built in TypeScript/Node.js (20+), Kryfto ships as a Docker Compose stack with a headless Chromium fleet, a FastAPI-compatible REST API, and a dedicated MCP server with 42 tools. The project targets developers who want full data ownership and no per-call vendor pricing.

From the README: “Self-host your own headless browser fleet. Connect it instantly to AI agents, IDEs, and workflow engines via OpenAPI and MCP.”

Key capabilities from the README:

• 42-tool MCP server for Claude Code, Cursor, and Codex — browse, search, extract, fact-check, and run continuous research agents
• Anti-bot stealth engine with 12 rotated User-Agents (Chrome 130–133, Edge 131/133), canvas fingerprint randomization, WebGL spoofing, WebRTC leak prevention, and humanized mouse movements
• CAPTCHA solving for Cloudflare Turnstile, reCAPTCHA v2, hCaptcha, and Datadom — using local CLIP vision and Whisper transcription (no external paid APIs)
• Admin dashboard on port 3001 for managing tokens, projects, jobs, audit logs, and rate limits, plus an API Playground
• Enterprise stack under the hood: Postgres for persistence, Redis + BullMQ for job queuing, MinIO/S3 for artifact storage

Setup Workflow

Prerequisites

• Node.js 20 or higher
• Docker and Docker Compose
• 4 GB RAM minimum (8 GB recommended for concurrent browser sessions)

Step 1: Clone and configure

git clone https://github.com/ExceptionRegret/Kryfto.git
cd Kryfto

# Auto-generate a secure .env with random tokens and passwords
node scripts/generate-env.mjs -o .env

# Or copy the example and fill in values manually
cp .env.example .env

Step 2: Launch the stack

# Spin up the entire infrastructure:
# API, Dashboard, Worker, Postgres, Redis, Minio S3
docker compose up -d --build

# Verify health
curl -H "Authorization: Bearer $KRYFTO_TOKEN" \
  http://localhost:8080/v1/healthz

The Admin Dashboard becomes available at http://localhost:3001/dashboard/ — log in with your admin API token.

Step 3: Connect via MCP

The MCP server exposes 42 tools. In Claude Code, Cursor, or Codex, point the MCP client to the Kryfto MCP endpoint:

# Example: dispatch a browsing job
curl -X POST http://localhost:8080/v1/jobs \
  -H "Authorization: Bearer $KRYFTO_TOKEN" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: demo-example-1" \
  -d '{"url":"https://example.com"}'

Step 4: Retrieve extracted content

# Fetch the extracted Markdown or HTML artifact
curl -H "Authorization: Bearer $KRYFTO_TOKEN" \
  http://localhost:8080/v1/jobs/<jobId>/artifacts

Step 5: Federated search

Kryfto natively queries DuckDuckGo, Brave, and Google in a single federated search call:

# Check the README for the current federated search endpoint
curl -H "Authorization: Bearer $KRYFTO_TOKEN" \
  http://localhost:8080/v1/search \
  -d '{"query": "your search term", "engines": ["duckduckgo", "brave", "google"]}'

Deeper Analysis

Architecture

Kryfto uses a per-domain browser session pool with a 30-minute TTL. Each session maintains its own cookie jar, user-agent, and fingerprint profile. This prevents cross-site cookie leakage and reduces bot detection from shared browser fingerprints.

The CAPTCHA solving pipeline is particularly notable: reCAPTCHA image grids are classified locally via CLIP vision (clip-vit-large-patch14), and audio challenges are transcribed locally via Whisper — both running through @xenova/transformers. No external paid service is required.

The job queue (Redis + BullMQ) ensures reliable concurrent job execution even if the API process restarts. Artifacts land in MinIO/S3 with permanent storage rather than ephemeral containerfs.

Self-hosting economics

The README claims “self-host for $5/mo flat.” A minimal setup on a budget VPS (2 vCPU, 4 GB RAM) can run the core stack. Production use with concurrent browser sessions and CAPTCHA solving benefits from 8 GB RAM and additional headless browser workers.

Comparison to alternatives

| Feature | Kryfto | Browserbase | Steel Thread | |---|---|---|---| | MCP server | 42 tools | Via API | No | | Self-hosted | Yes | No (cloud only) | No | | CAPTCHA solving | Local CLIP/Whisper | External API | None | | Anti-bot stealth | 12 UAs, canvas spoofing | Yes | Basic | | License | Apache 2.0 | Proprietary | Open-source | | Docker-based | Yes | No | Varies |

Practical Evaluation Checklist

• Clone and run docker compose up locally
• Verify /v1/healthz returns 200 with the API token
• Open the Admin Dashboard at http://localhost:3001
• Dispatch a test extraction job via cURL
• Verify artifact retrieval with /v1/jobs/<id>/artifacts
• Test MCP tool access from Claude Code or Cursor (if configured)
• Check the SLO dashboard for per-tool success rates

Security Notes

• All HTTP extractions execute purely in-memory with no telemetry persisted to Postgres
• API authentication uses Bearer tokens (auto-generated via generate-env.mjs)
• Per-role rate limits are configurable in the Admin Dashboard
• WebRTC IP leak prevention blocks ICE candidate leaks in browser sessions
• Idempotency-Key header on job creation prevents duplicate submissions

FAQ

Q: Does Kryfto require a GPU? A: No. The core extraction pipeline runs on CPU. GPU acceleration is only needed if you enable the local CLIP/Whisper CAPTCHA solving pipeline — and even then, it falls back gracefully to non-GPU execution.

Q: How does it differ from Playwright or Puppeteer directly? A: Playwright and Puppeteer are browser automation libraries. Kryfto wraps them with an MCP server interface, anti-bot evasion (rotated UAs, canvas spoofing, humanized mouse curves), CAPTCHA solving, federated search across multiple engines, and a production job queue — all behind a REST API with an admin UI.

Q: Can I use Kryfto with n8n or Zapier? A: Yes. Kryfto ships with a fully documented OpenAPI spec and a REST API, making it straightforward to drop into n8n, Zapier, Make, or any custom TypeScript/Python pipeline.

Q: What is the minimum VPS spec for self-hosting? A: The README suggests $5/month for a minimal setup (2 vCPU, 4 GB RAM). Production use with concurrent headless browsers and CAPTCHA solving benefits from 8 GB RAM.

Q: Is there a managed cloud version? A: No — Kryfto is fully self-hosted only. There is no official hosted or managed offering.

Conclusion

Kryfto bridges the gap between static LLM knowledge and live web interaction. Its 42-tool MCP server gives AI agents real browsing, search, and extraction capabilities without sending data to a third-party scraping service. The anti-bot stealth engine and local CAPTCHA solving address the two biggest blockers for production web automation.

If you need to give Claude Code, Cursor, or Codex persistent access to dynamic, bot-protected web content — and you want to self-host for full data ownership — Kryfto is worth a closer look.

Get started at github.com/ExceptionRegret/Kryfto.