Fly.io Reference

Commands, customization, environment variables, and troubleshooting for the Fly.io template.

Fly app names are global, so up.sh generates a unique name (agentos-<suffix>), records it in fly.toml, and pairs it with a Postgres app named <app>-db. Later fly commands read the app name from fly.toml, so no --app flag is needed.

Manage

Task	Command
Deploy code changes	./scripts/fly/redeploy.sh
Sync env variables	./scripts/fly/env-sync.sh (defaults to .env.production; pass .env to sync that instead)
Tail logs	fly logs
Resize the machine	Edit [[vm]] in fly.toml, then ./scripts/fly/redeploy.sh
Tear down	./scripts/fly/down.sh (add --yes to skip the confirmation)

The platform runs one machine by design. Both deploy scripts pass fly deploy --ha=false because the Fly default creates two machines, which doubles cost and runs two in-process schedulers double-firing every cron. fly.toml keeps the machine warm with auto_stop_machines = "off" and min_machines_running = 1; with scale-to-zero, scheduled jobs silently stop firing.

Production auth

Token-Based Authorization is on by default. Production startup requires JWT_VERIFICATION_KEY or a readable JWKS file at the container path in JWT_JWKS_FILE; otherwise the process exits.

Token-Based Auth gives you three things:

1. No public access. The server rejects requests without a valid token.
2. Per-request identity. Middleware validates the token and exposes its user_id, optional session_id, scopes, and claims to the request.
3. Scope-based permissions. Token scopes control access to AgentOS routes and resources.

The templates do not enable per-user data isolation. To scope non-admin session, memory, trace, and run access to the JWT subject, pass authorization_config=AuthorizationConfig(user_isolation=True) to AgentOS. See User Isolation.

To opt out (not recommended), set authorization=False in app/main.py and redeploy. Use this only inside a private network behind another auth layer. Without it, anyone who guesses your Fly URL can access your platform.

Customize

Add an agent

Ask your coding agent to run /create-agent, or do it by hand. Create agents/my_agent.py:

from agno.agent import Agent

from app.settings import default_model
from db import get_postgres_db

INSTRUCTIONS = """\
What the agent does, which tools it uses, the rules to follow when answering.
"""

my_agent = Agent(
    id="my-agent",
    name="My Agent",
    model=default_model(),
    db=get_postgres_db(),
    instructions=INSTRUCTIONS,
    enable_agentic_memory=True,
    add_datetime_to_context=True,
    add_history_to_context=True,
    num_history_runs=5,
)

Register it in app/main.py:

from agents.my_agent import my_agent

agent_os = AgentOS(
    ...
    agents=[agent_builder, platform_manager, web_search, my_agent],
)

Add its UI metadata beneath the existing manifest: key in app/config.yaml:

  my-agent:
    description: "What the agent does."
    quick_prompts:
      - "First example prompt"
      - "Second example prompt"
      - "Third example prompt"

Local containers hot-reload on save. For production, run ./scripts/fly/redeploy.sh.

Change the model

app/settings.py defines default_model(), used by every agent. Change it in one place:

from agno.models.anthropic import Claude

def default_model():
    return Claude(id="claude-sonnet-5")

Add anthropic to pyproject.toml, set the provider key in your env, and regenerate pins:

./scripts/generate_requirements.sh

Rebuild locally with docker compose up -d --build. For production:

./scripts/fly/env-sync.sh
./scripts/fly/redeploy.sh

Add tools

Agno ships 100+ toolkits. See Toolkits.

from agno.tools.slack import SlackTools

my_agent = Agent(
    ...
    tools=[SlackTools()],
)

Add dependencies

1. Edit pyproject.toml.
2. Regenerate pins: ./scripts/generate_requirements.sh (add upgrade to refresh every pin).
3. Rebuild locally with docker compose up -d --build, or redeploy with ./scripts/fly/redeploy.sh.

Enable Slack

Set both variables in your env file:

SLACK_BOT_TOKEN=xoxb-...
SLACK_SIGNING_SECRET=...

Sync with ./scripts/fly/env-sync.sh. The interface activates automatically and routes messages to Agent Builder; change the agent= argument in app/main.py to point at another agent. See Slack setup.

Toggle scheduled workflows

The deployment check runs daily by default (ENABLE_DEPLOY_CHECK=True); it is deterministic and free. The run-evals schedule is always registered but starts disabled because it uses model calls. Enable it from the AgentOS UI. Both workflows remain runnable on demand.

Enable pgvector for knowledge bases

Fly’s stock postgres-flex image does not ship pgvector, so sessions and memory work but knowledge bases (RAG) fail at CREATE EXTENSION time. The fix is a two-line Dockerfile:

FROM flyio/postgres-flex:17
RUN apt-get install -y postgresql-17-pgvector

Set FLY_PG_IMAGE to that image before running ./scripts/fly/up.sh. The image applies when the Postgres cluster is created; up.sh reuses an existing cluster, so switching later means tearing down with ./scripts/fly/down.sh and provisioning fresh.

Format, validate, and run evals

The format, validate, and eval scripts run on the host and need a venv. Set it up once:

./scripts/venv_setup.sh
source .venv/bin/activate

Task	Command
Format	./scripts/format.sh
Lint and type-check	./scripts/validate.sh
Run smoke evals	python -m evals --tag smoke

./scripts/mcp_check.sh runs inside the container, so it needs no venv.

Environment variables

Variable	Required	Default	Description
OPENAI_API_KEY	Yes	-	Models and embeddings.
RUNTIME_ENV	No	prd	dev disables JWT. Compose sets it for local. Never put it in an env file that syncs to Fly, or production deploys unauthenticated.
JWT_VERIFICATION_KEY	Production	-	Public key from os.agno.com. Quote the value so the multi-line PEM parses as one variable.
JWT_JWKS_FILE	Production	-	Path inside the Fly Machine to a JWKS file. The Fly scripts sync only this path. Bake the file into the image or configure a Fly mount before deploying.
MCP_CONNECT_SECRET	No	generated by up.sh	OAuth consent secret (16+ chars) for connecting claude.ai and ChatGPT to /mcp. up.sh generates one on deploy and writes it to .env.production.
AGENTOS_MCP_SIGNING_KEY	No	generated	Optional high-entropy signing-key material (32+ chars) for OAuth tokens. Unset, a strong key is generated and persisted in the database. Rotating it invalidates outstanding tokens.
AGENTOS_URL	No	http://127.0.0.1:8000	Scheduler base URL. up.sh sets it to https://<app>.fly.dev. Scheduled jobs never fire if it stays at the default in production. Re-running up.sh resets a hand-set value to the generated fly.dev URL, so re-pin the domain (or re-run env-sync.sh) afterwards. When MCP_CONNECT_SECRET is set, OAuth metadata also derives its public origin from this URL.
ENABLE_DEPLOY_CHECK	No	True	Daily deployment-check cron.
EVALS_TAG	No	smoke	Eval tag the run-evals workflow runs.
EVALS_CASE_TIMEOUT_SECONDS	No	90	Per-case timeout for run-evals runs.
EVALS_SUITE_TIMEOUT_SECONDS	No	900	Whole-suite timeout for run-evals runs.
PARALLEL_API_KEY	No	-	WebSearch uses the Parallel SDK when set, keyless MCP otherwise.
SLACK_BOT_TOKEN	No	-	Set with the signing secret to enable Slack.
SLACK_SIGNING_SECRET	No	-	Set with the bot token to enable Slack.
DB_HOST / DB_PORT / DB_USER / DB_PASS / DB_DATABASE	No	matches compose	Postgres connection. up.sh sets these as Fly secrets pointing at <app>-db.flycast.
DB_DRIVER	No	postgresql+psycopg	SQLAlchemy driver.
AGNO_DEBUG	No	False	Verbose Agno logs. Compose sets it for dev.
WAIT_FOR_DB	No	False	If True, the entrypoint blocks on the database before starting. Compose sets it.

up.sh also reads three variables from your shell when provisioning:

Variable	Default	Description
FLY_REGION	iad	Region for the app and its Postgres. Re-runs keep the region already in fly.toml unless you set this explicitly.
FLY_ORG	personal	Fly org for both apps. They must share an org or the private network between them does not exist.
FLY_PG_IMAGE	stock postgres-flex	Postgres image. Point it at a pgvector-enabled derivative to support knowledge bases.

Troubleshooting

flyctl: command not found

Install flyctl, then run fly auth login. The scripts accept either binary name, flyctl or fly.

up.sh pauses asking for a JWT key

Expected. At os.agno.com, choose Connect OS → Live, enter your Fly URL, name it Live AgentOS, turn on Token-Based Authorization (JWT) on the connection panel, and connect. The UI generates the public key. If the OS is already connected, enable the setting under Settings → OS & Security. Paste the full PEM into the script prompt. To add a PEM later, set JWT_VERIFICATION_KEY and run ./scripts/fly/env-sync.sh. To use JWKS, bake or mount the file in the Fly Machine, set JWT_JWKS_FILE to its container path, then sync the variable.

up.sh exits: fly.toml carries an app name this script doesn't manage

The script only manages names it generated. A different name means continuing would overwrite fly.toml and abandon that app. Restore the agentos placeholder (or an agentos-* name from a previous run), or tear down first with ./scripts/fly/down.sh.

App fails to start in production

JWT auth is on whenever RUNTIME_ENV is not dev. Set JWT_VERIFICATION_KEY and sync. To use JWT_JWKS_FILE, first bake or mount a readable JWKS file at that container path, then sync the variable. env-sync.sh does not upload the file. To opt out inside a private network behind another auth layer, set authorization=False in app/main.py.

Knowledge bases fail at CREATE EXTENSION

The stock postgres-flex image has no pgvector; sessions and memory are unaffected. Set FLY_PG_IMAGE to a pgvector-enabled image and recreate the cluster. up.sh reuses an existing cluster, so tear down first with ./scripts/fly/down.sh. See Enable pgvector for knowledge bases.

Scheduled jobs never fire

AGENTOS_URL may still be the localhost default. up.sh sets it to https://<app>.fly.dev automatically; for a custom domain or tunnel, set it by hand and run ./scripts/fly/env-sync.sh. Re-running up.sh resets a hand-set value to the generated fly.dev URL, so re-pin it afterwards. The other cause is the machine scaling to zero: keep auto_stop_machines = "off" and min_machines_running = 1 in fly.toml.

Every cron fires twice

A plain fly deploy created a second machine, and each runs its own scheduler. Deploy with the scripts; they pass --ha=false.