SSD Nodes Learn Hosting plans →
How to do am Matt ConnorBy Matt Connor

Best self-hosted Firecrawl alternatives for VPS

Compare Draco, Hound, and self-hosted Firecrawl for your VPS. We check RAM usage, headless browser needs, and API compatibility. Get the pinned install and MCP setup.

Wetin self-hosted Firecrawl alternative suppose do

Self-hosted Firecrawl alternative get one main work: make e take URL and return the page as clean markdown wey agent fit read. Hosted APIs dey charge for each page, so bill go increase as your agent dey curious. VPS wey you already dey pay for fit do the same work. The projects differ for one question: headless browser (real browser engine wey dey run without window) need start for your machine?

This answer determine memory footprint, cost for each page, and which pages go return empty. This guide compare Draco, Hound and self-hosted Firecrawl release. E install the lightest one for pinned version, then connect am to agent through MCP (model context protocol).

Di four projects, and wetin each one be

Draco na one binary, wey dem write for Rust, and e get MIT or Apache-2.0 license. draco scrape <url> dey print markdown go stdout. draco serve dey run one daemon wey dey answer for 127.0.0.1:3002, di port wey Firecrawl dey use. As dem document am for 12 October 2026, e no get container image and e no dey start any browser by default. As of 12 October 2026, di upstream repository for github.com/0xchasercat/draco, di account wey get am, and di release downloads all dey return 404 Not Found. As of dat check, one public fork still get di source and tags up to v0.24.0 but e no publish any release builds, so read di install section wey dey down here before you plan your work around am.

Firecrawl self-hosted na di engine wey dey behind di hosted product, under AGPL-3.0. Im docker-compose.yaml define seven services: playwright-service, api, redis, rabbitmq, nuq-postgres, foundationdb and foundationdb-init. You go get di real crawl queue, but you go pay di price of running one small distributed system.

Hound dey live for di master-fetch repository and e dey ship go PyPI as hound-mcp, e get MIT license, version 13.2.0 (wey dem publish for 26 August 2026) as of 12 October 2026. E need Python 3.11 or newer. E be MCP server first and fetcher second: e dey try plain HTTP and e go upgrade go Patchright browser if di plain fetch come back blocked or e serve JavaScript shell; di browser-enabled setup dey warm one browser wen e start.

Trawl dey here because pipo dey see am wen dem dey search for di oda ones, and e dey do different work. E dey solve JavaScript challenges and CAPTCHAs wit one fingerprint-patched Firefox, as replacement for FlareSolverr for one *arr media stack. Now, e dey read Markdown and e dey extract structured JSON over MCP. Di section on etiquette wey dey down here explain why di challenge-solving role wey e get go decide weda e fit join your agent stack at all.

Why browser pool dey make small VPS boxes crash

Each browser tab wey dey open na separate renderer process wey dey hold its own DOM (document object model) and its own JavaScript heap. So memory dey increase based on how many pages dey open at the same time, not how many pages you fetch per day. Two of these projects write this cost for their own compose files.

ChartMemory ceilings each project sets in its own compose file (GB)
Di data wey dey behind dis chart
[
  {
    "label": "Firecrawl api",
    "memory_limit_gb": 8
  },
  {
    "label": "Firecrawl playwright",
    "memory_limit_gb": 4
  },
  {
    "label": "Hound (browser included)",
    "memory_limit_gb": 3
  }
]

The Firecrawl compose file limit its api container to 8 GB and its Playwright container to 4 GB, with matching swap limits. Hound's compose set 3 GB for one container wey carry bundled Chromium. These na limits wey the projects choose and publish, not measurements from an idle system. Redis, RabbitMQ, PostgreSQL and FoundationDB still need their own share on top of the Firecrawl figures.

A limit wey pass the RAM wey you get no reserve that memory. When the box run out of memory, the kernel out-of-memory killer go terminate one process. So container fit disappear from docker compose ps without any error for the application log. Firecrawl's supplied Compose file no set automatic restart policy, so container wey OOM terminate go remain stopped; docker compose ps --all include stopped containers. Read dmesg -T | tail after any unexplained stop. Firecrawl no publish any verified minimum host size for this stack, so size the box based on your workload. Memory limits for Docker Compose explain how to set the numbers for each service.

One more browser detail fit waste person whole evening. Docker give container 64 MB shared memory at /dev/shm, but Firecrawl's Playwright service start Chromium with --disable-dev-shm-usage. This one avoid that small partition for these shared-memory files. Hound's compose increase the size with shm_size: "1gb". Its dynamic browser tier inherit --disable-dev-shm-usage from Playwright through Patchright, and neither Patchright's switch removals nor Hound's ignore_default_args list remove am. Firecrawl's compose no set shm_size at all. Its playwright-service block set CPU and memory limits and 1 GB tmpfs at /tmp/.cache, so that container keep the 64 MB default. Check the Chromium launch flags before you copy Hound's line into an image wey you build around Playwright.

Extraction quality on JavaScript-heavy pages

For static HTML, server-rendered blog, documentation page, or news article, dem all dey return almost the same markdown, so the fastest one wins. Difference dey show for client-rendered pages, where delivered HTML na empty shell and JavaScript go bring the text after page load.

Draco dey escalate by tiers. Tier 0 and tier 1 parse HTML without any JavaScript. Tier 2 dey run the page own JavaScript inside an in-process V8 isolate. Na the JavaScript engine be this, but browser no dey around am. README talk say page code no get host capability bindings for there. This cover plenty single-page applications with small part of browser memory. When Draco meet wall wey e no fit pass, draco scrape go exit with code 3, needs_browser. Check this for scripts, because empty file with zero exit code na the failure wey quietly spoil agent context:

draco scrape https://example.com > page.md
echo "exit=$?"

Firecrawl's playwright-service dey control real Chromium, so e render wetin browser dey render. But self-hosted build still no be the hosted product. Documentation talk say Fire-engine, wey carry cloud service advanced anti-bot behaviour, no dey included. E also talk say screenshots and page actions no dey available for default stack. For Agent, Browser and interact, e direct you to Firecrawl Cloud. Hound dey between both options intentionally. E fetch over HTTP and escalate for each request. Its warm browser go close after idle timeout, so quiet box go remain near baseline.

Draco install with pinned version

README document one one-line installer. Read wetin e dey do before you pipe am go shell: as dem document am on 12 October 2026, e dey install go $HOME/.draco/bin/draco, e always take latest release, and e no check any signature or hash. For server, pin the version and verify the download. The v0.20.5 instructions below still get the setup we check on 12 October 2026; upstream documentation for the pinned version no dey available to check again.

As of 12 October 2026, the project's release downloads dey return 404 Not Found because upstream repository no longer dey answer, and the get.draco.sh installer domain wey the project's site advertise no resolve. No live release URL dey wey you fit fetch from, so the commands below start for a machine wey you already get the v0.20.5 archive, draco-linux-x86-64.tar.gz, and its SHA256SUMS for /tmp. You fit don keep dem from earlier download or copy dem from machine wey you trust. No use mirror wey you never vet to fill the gap: checksum file wey the same mirror serve no prove anything about the binary.

cd /tmp
sha256sum --ignore-missing -c SHA256SUMS

That one go print draco-linux-x86-64.tar.gz: OK. If you see a FAILED line, e mean say the bytes wey you hold no be the same bytes wey project publish, so delete dem and start again.

mkdir -p draco-v0.20.5
tar -xzf draco-linux-x86-64.tar.gz -C draco-v0.20.5
sudo install -m 755 "$(find draco-v0.20.5 -type f -name draco | head -n1)" /usr/local/bin/draco
draco scrape https://example.com

The last command go print the example page as markdown. The find no be decoration: archive layout no dey part of the project's public contract, and the official installer wey we last check on 12 October 2026 dey locate the binary the same way.

Run the daemon with its own account instead of your login user. Write /etc/systemd/system/draco.service:

[Unit]
Description=Draco fetch daemon
After=network-online.target
Wants=network-online.target

[Service]
User=draco
ExecStart=/usr/local/bin/draco serve --host 127.0.0.1 --port 3002 --max-concurrency 4
Restart=on-failure
NoNewPrivileges=true
ProtectSystem=strict
ProtectHome=true
PrivateTmp=true

[Install]
WantedBy=multi-user.target
sudo useradd --system --no-create-home --shell /usr/sbin/nologin draco
sudo systemctl daemon-reload
sudo systemctl enable --now draco
curl -s http://127.0.0.1:3002/health

/health go answer as soon as the daemon dey listen. Connection refused mean say e no dey listen, so read journalctl -u draco -n 50. The usual cause na another process wey already dey hold 3002, because na also Firecrawl default port. --port go move Draco. Self-hosted Firecrawl no get that kind flag: its port dey come from the PORT variable for its Compose file, published as "${PORT:-3002}:${INTERNAL_PORT:-3002}", so change the published port there and make every client use the same one. Unit files in more detail: systemd service units and timers.

Now fetch am the same way your agent go:

curl -X POST http://127.0.0.1:3002/v1/scrape \
  -H 'content-type: application/json' \
  -d '{"url": "https://example.com", "formats": ["markdown"]}'

Keep fetch daemon comot public internet

Fetch API wey no get authentication na open proxy. Anybody wey fit reach the port fit make your server request any URL under your IP address, and abuse report go go your provider, no be to the person wey cause am. Draco's serve flags wey documentation list on 12 October 2026 no include API key, so na network suppose provide the protection. Keep default 127.0.0.1 bind when agent dey run for the same box. When agent dey another place, put both ends for private tunnel; WireGuard VPN wey you host by yourself na the usual answer. Bind to the tunnel address instead of 0.0.0.0. Then check from another machine say public IP no dey answer anything. ufw firewall basics and least privilege user accounts cover the two sides of this.

Agent code bilong go change? API compatibility for real life

Draco dey answer Firecrawl v1 routes: /v1/scrape, /v1/map, /v1/crawl, /v1/batch/scrape and /v1/search. E README, wey dem last check on 12 October 2026, talk say e dey accept unknown fields and ignore dem. Compatibility depend on the features wey you request. Draco official site talk say screenshots and clicks dey return HTTP 422. So, Firecrawl v1 client wey request unsupported page actions need changes for request and new base URL too. Watch wetin move for the other side. Firecrawl own self-hosting guide now dey smoke-test with POST to /v2/scrape, and current SDKs dey use v2 by default, while Draco site list v1 routes. Test every call with curl before you edit agent code. Read the JSON body instead of only the status code, because na field names dey different between these implementations.

Robots.txt, rate limits, and the line to stay behind

Draco dey read robots.txt by default, and e README wey dem last check on 12 October 2026 document --ignore-robots to turn am off. Firecrawl document the same default. Leave both as dem dey. Hound dey check robots.txt only when respect_robots=True pass to smart_fetch, so enable am there. Then set your own pace: as dem document on 12 October 2026, Draco --delay dey put milliseconds between requests, while --max-concurrency dey limit parallel jobs, with 8 as the daemon default. Two to four dey gentler on shared VPS link and e rarely slower overall, because site wey start rate limiting you go cost more minutes than the concurrency wey you save. Cache wetin you fetch, so second agent run no go cost the source anything. That one still be the cheapest line item for controlling wetin AI agent dey cost you.

Challenge walls na separate matter, and Trawl dey built exactly for that: Cloudflare Turnstile, reCAPTCHA, hCaptcha and GeeTest. Challenge wall na when site clearly refuse automated traffic. To work around am go put you against the site's terms of use, and for some places against the law, so this guide cover fetching infrastructure and stop there. The same techniques wey fit pass wall na the ones site owners dey watch and block, so any pipeline wey depend on dem go fragile and rude. When source important reach that level, look for its RSS feed, public API, or bulk export. Each one cheaper to run, and none go break the week the wall change.

Connect am go agent via MCP

MCP (model context protocol) na di interface wey agent dey use call tool. Na di harness wey di model dey run inside dey decide which tools di model fit reach, and weda e fit call dem without ask you first. So, to make di daemon dey listen na just half of di work. Draco get MCP server inside di same binary, wey dey use stdio:

{ "mcpServers": { "draco": { "command": "draco", "args": ["mcp"] } } }

Di README wey dem check last for 12 October 2026 list draco_scrape, draco_discover and draco_search for stdio; di draco_interact_* set na for di daemon HTTP MCP endpoint. Stdio only dey work if di agent process and di binary dey di same machine, sake of say di transport na di process standard input. For agent wey dey another host, Hound dey serve MCP via HTTP instead: hound --http --host 127.0.0.1 --port 8765 dey publish endpoint for http://127.0.0.1:8765/mcp, wey you go reach across di tunnel. You go see transport choices and wetin you fit expose for running MCP servers on a VPS.

Fetch pairs wit search. Agent wey fit only fetch go wait make you give am URLs. Add a self-hosted SearXNG search instance so e fit find dem by imsef, di same way as browser search skill built on SearXNG. Once di daemon don start, e be one shared service for any of di self-hosted AI agents wey you run. If di agent side be di part wey you no too sure of, writing a small agent loop by hand first go make am clear wia fetch tool dey plug in and wetin di model dey do wit di markdown wey e collect.

FAQ

I need headless browser to fetch pages for AI agent?

No be for most pages. Documentation, blogs, and news articles wey server render dey come back complete with ordinary HTTP fetch plus HTML-to-markdown step. Na this method Draco dey use for lower tiers, and project figures talk say e fit take about 300 ms per page without browser. Browser dey useful for client-rendered applications, where HTML wey server deliver na empty shell. Draco V8 isolate fit handle plenty of this middle ground without browser process. If e no fit, e dey exit with code 3, needs_browser.

How much RAM self-hosted Firecrawl need for VPS?

Its compose file set 8 GB limit for api container and 4 GB for Playwright container. The same stack also start Redis, RabbitMQ, PostgreSQL, and FoundationDB. Firecrawl never publish any verified minimum host size for this stack. When memory pressure happen, kernel out-of-memory killer go terminate processes. If container stop, the supplied Compose file no go restart am automatically. Check docker compose ps --all for stopped containers and use dmesg -T | tail confirm say OOM event happen.

Draco be drop-in replacement for Firecrawl API?

For v1 endpoints, e dey close. E serve /v1/scrape, /v1/map, /v1/crawl, /v1/batch/scrape, and /v1/search. But compatibility depend on the features wey request use. Screenshots and clicks return HTTP 422, so clients wey request unsupported page actions need change their requests and use new base URL. Firecrawl current SDKs dey use v2 by default, while Draco site list v1 routes. Verify every call wey your agent make with curl first.

If I self-host scraper, e mean say I fit ignore robots.txt?

No. Where the code dey run no change wetin website publish or wetin its terms allow. Draco and Firecrawl both respect robots.txt by default. The override flag dey for sites wey you own or wey give you written permission to crawl. Rate limits still dey enforced for the other end, so polite --delay with low concurrency go help your IP address keep working. If a stack only work by defeating challenge wall, e fit stop working without warning.

Dis article help you?

#scraping#firecrawl#ai-agents#self-hosting#markdown