minwebrender

Headless-Chromium page renderer. Given a URL it loads the page in Chromium, strips scripts/styles/images, and returns the readable text as markdown.

Its consumer is the web_render agent tool in safegraph-agentic-research, which calls it for every page an agent needs to read. web_render falls back to Firecrawl, then to Serper's scrape API, when this service is unreachable.

Endpoints

Route	Purpose
GET /<url>	Render as an HTML page. The contract web_render speaks today.
GET /text/<url>	Same render as text/plain, skipping the markdown→HTML round trip.
GET /healthz	Liveness. 503 when Chromium is not connected.
GET /readyz	Readiness. Same signal.
GET /	Human-facing form.

Configuration

Env var	Default	Purpose
DOMAIN	0.0.0.0:10000	Host stamped onto rewritten <a href>s. Set to the Service address.
MAX_PAGES	6	Concurrent Chromium pages. Drives the memory limit.
PAGE_TIMEOUT_MS	45000	Per-page navigation budget. Keep under the caller's tool timeout.
REQUEST_TIMEOUT_MS	75000	Ceiling on the whole request, including the wait for a free page slot. Keep under gunicorn's --timeout.
PORT	10000	Dev server only; gunicorn binds 10000 in the image.
LOG_LEVEL	INFO

Local development

pip install -r requirements.txt && playwright install chromium
python main.py                        # http://localhost:10000
curl localhost:10000/text/example.com

Or via the image, which is what CI and EKS run:

docker build -t minwebrender . && docker run --rm -p 10000:10000 minwebrender

Deployment

Runs on SafeGraph's EKS cluster, built and shipped the same way as safegraph-agentic-research:

GitLab CI (.gitlab-ci.yml)          Harness
build -> test -> push to ECR   -->  helm release `minwebrender`
                                    (deploy/helm/)

• Image: $ECR_REGISTRY/minwebrender, tag latest from the default branch. Built on python:3.10-bullseye and installs Chromium with playwright install --with-deps, so the browser revision is chosen by the pinned playwright package. Debian 11 is deliberate: glibc 2.34+ calls clone3, which the CI runner's Docker 19.03 rejects, and no thread can then start. See the comment at the top of the Dockerfile.
• Values: deploy/helm/ — base.yaml plus values-production.yaml (3 replicas) and values-staging.yaml (1 replica).
• Internal to the VPC: the shared chart has no ClusterIP-only mode, so a subdomain is set and an internal ALB is created. loadbalancerIsPublic stays false, so nothing is reachable from the internet — the old public minwebrender.net was an unauthenticated open renderer where anyone could make SafeGraph infrastructure fetch any URL.
• In-cluster callers use http://minwebrender and never touch the load balancer. Port 80, not 10000: the chart puts the Service on 80 and forwards to the container's 10000.
• No AWS permissions: the renderer touches no S3, RDS, or Secrets Manager, so it needs no IRSA role and no secrets.
• CI test: boots the built image and renders a page end-to-end. That is the check that catches Chromium/playwright version skew and a browser that will not launch under the container's seccomp profile.
• One gunicorn worker on purpose: one worker means one Chromium. A second worker process would mean a second browser and double the memory for no extra throughput. Scale with replicas.

Previously deployed on Render.com behind minwebrender.net.

Differences from the Render deployment

• _ensure_browser() relaunches a disconnected browser under a lock. The old code launched Chromium once at import and never re-checked it, so a dead browser wedged the process until someone redeployed — the cause of the 2026-09-03 outage, when the single instance stopped answering after a handful of concurrent renders and did not recover.
• /healthz and /readyz, so the kubelet can restart what a relaunch cannot fix, and 3 replicas so there is no single instance to wedge.
• --no-sandbox --disable-dev-shm-usage, needed under a pod's seccomp profile and 64Mi /dev/shm.
• Not exposed to the internet.