Skip to content
Deploying Agentic AI

Render

The primary deployment path for this project. Free web services, no credit card, deploys straight from a GitHub repo.

Two ways to do it

Section titled “Two ways to do it”

The repo already contains deploy/render.yaml, so the settings are in version control rather than in someone’s memory.

  1. Render → New + → Blueprint
  2. Pick your repo
  3. Render reads the file and asks you for the three secret values
  4. Apply

The blueprint, explained

Section titled “The blueprint, explained”
deploy/render.yaml
services:
  - type: web
    name: research-agent
    runtime: python
    plan: free
    rootDir: app
    buildCommand: pip install uv && uv sync --frozen
    startCommand: uv run fastapi run main.py --port $PORT
    healthCheckPath: /health


    envVars:
      - key: LLM_API_KEY
        sync: false
      - key: SUPABASE_URL
        sync: false
      - key: SUPABASE_SECRET_KEY
        sync: false
      - key: TAVILY_API_KEY
        sync: false


      - key: LLM_MODEL
        value: gemini-flash-latest
      - key: LLM_PROVIDER
        value: google
      - key: MAX_AGENT_STEPS
        value: "12"
      - key: MAX_THREAD_TURNS
        value: "10"
      - key: ALLOWED_ORIGINS
        value: "*"

Four things in there are worth knowing:

  • rootDir: app points Render at the uv project — pyproject.toml, uv.lock, and the app code all live in app/, not the repo root. Every other path in the blueprint, including buildCommand and startCommand, resolves relative to it.
  • uv sync --frozen means “use uv.lock exactly, don’t re-resolve”. Your deploy gets the versions you tested, not whatever is newest today.
  • sync: false means “ask me for this value in the dashboard, and never store it in the repo”. This is how secrets stay out of git — the four keys above it are secret, so none of them carry a value here. TAVILY_API_KEY is the one that’s also optional: leave it empty in the dashboard and the agent runs on Wikipedia alone.
  • The variables that do carry a value (LLM_MODEL, LLM_PROVIDER, MAX_AGENT_STEPS, MAX_THREAD_TURNS, ALLOWED_ORIGINS) aren’t secret, so committing them is fine. Change ALLOWED_ORIGINS to your real frontend origin before you go public.

Environment variables

Section titled “Environment variables”

Dashboard → your service → Environment.

Variable Required Notes
LLM_API_KEY yes from AI Studio, or whichever provider you chose
SUPABASE_URL yes no trailing slash
SUPABASE_SECRET_KEY yes the sb_secret_… key, not the publishable one. On a pre-rename project, the legacy service_role key works here too
LLM_MODEL on Google only defaults to gemini-flash-latest; required for every other provider
LLM_PROVIDER no google (default), cerebras, openrouter, openai-compatible
LLM_BASE_URL only for openai-compatible e.g. https://api.groq.com/openai/v1
TAVILY_API_KEY no enables web search, so the agent can answer questions about recent events
MAX_AGENT_STEPS no defaults to 12
MAX_THREAD_TURNS no defaults to 10 — the cap on one conversation
ALLOWED_ORIGINS no comma-separated, no trailing slash

The old GEMINI_API_KEY and GEMINI_MODEL names still work when LLM_PROVIDER is google, so a service deployed before this change keeps running untouched.

What the free tier actually gives you

Section titled “What the free tier actually gives you”
Instance hours 750 / month / workspace
Credit card not required
Spin-down after ~15 minutes idle
Cold start ~1 minute
CPU a small shared slice
Filesystem ephemeral — wiped on every restart

Two of these bite people:

The cold start. Hit /health ten minutes before any demo. A judge or interviewer waiting 60 seconds on a blank page will assume it’s broken.

The ephemeral filesystem. Anything written to disk is gone on restart. This is why the project keeps everything in Postgres and nothing on disk — it’s not an accident.

The small CPU slice is fine here because agent work is I/O-bound; you’re waiting on the model anyway. It would be bad for anything CPU-heavy.

Reading the deploy log

Section titled “Reading the deploy log”

Render shows you the build in real time. Worth knowing what you’re looking at:

==> Cloning from https://github.com/...
==> Running build command 'pip install uv && uv sync --frozen'...
    Resolved 94 packages in 12ms
    Installed 94 packages in 1.2s
==> Build successful 🎉
==> Deploying...
==> Running 'uv run fastapi run main.py --port $PORT'
    INFO:     Uvicorn running on http://0.0.0.0:10000
==> Your service is live 🎉

If it stops after “Deploying…” and eventually says something about ports, your start command isn’t binding 0.0.0.0.

Keeping it warm

Section titled “Keeping it warm”

The free tier sleeps. If you need it responsive:

  • Hit /health before demos, manually.
  • Or schedule a GitHub Action to ping it. Note this consumes your 750 monthly instance hours faster — pinging every 10 minutes keeps it awake permanently, which is roughly 720 hours a month. That’s within budget for one service, and over it if you run two.

Don’t ping more often than you need to.

See also

Section titled “See also”