Environment and Secrets Management Across Deployment Targets

Agent deployments fail loudly when keys are missing and fail expensively when keys leak or are shared without caps.

Use this cheatsheet to keep config, secrets, and environments consistent across laptop, serverless, containers, and VPS.

How to Use This Cheatsheet

Treat every row as a gate before promoting a release.
Separate non-secret config (model slug, log level) from secrets (API keys, DB passwords).
Prefer one naming scheme (AGENT_MODEL, OPENROUTER_API_KEY) across all hosts.
Re-audit after adding a new tool that needs credentials.

A - Secret Classes

Class	Examples	Handling
Model provider keys	OpenRouter, OpenAI, Anthropic	Secret store; per-env keys; spend caps
Tool credentials	GitHub PAT, Slack bot, Stripe	Least privilege; scoped to tool identity
Data stores	DB URL, Redis URL	Secret; rotate with app deploy
Signing / session	JWT secret, webhook HMAC	Secret; never log
Non-secret config	`AGENT_MODEL`, `LOG_LEVEL`, `PORT`	Plain env or config map OK

B - Injection by Target

Target	Prefer	Avoid
Local dev	`.env` (gitignored) + direnv/dotenv	Committing `.env`
Docker CLI	`-e`, `--env-file`, orchestrator secrets	`ENV KEY=secret` in Dockerfile
Compose	`${VAR}` from host env or secrets	Hard-coded keys in compose committed files
Vercel	Project env (Production / Preview / Development)	Same key for preview + prod
AWS Lambda	Function env + Secrets Manager/SSM	Long-lived keys when IAM role works
Kubernetes	Secret → env or volume mount	Base64 Secrets without encryption at rest / RBAC
VPS systemd	`EnvironmentFile=` mode `600`	World-readable unit drop-ins with keys
CI	OIDC to cloud / short-lived tokens	Long-lived cloud keys in plain CI vars when avoidable

C - Environment Matrix

Variable	Local	Preview/Staging	Production
`OPENROUTER_API_KEY`	Dev-capped key	Staging key	Prod key
`AGENT_MODEL`	Cheap default	Prod-like	Prod pin
`LOG_LEVEL`	debug	info	info/warn
`ENV` / `APP_ENV`	development	staging	production
Tool write credentials	Mock or sandbox	Sandbox	Prod least privilege

Never let staging share production write credentials "for convenience."

D - Naming and Loading Pattern

# pattern: explicit required secrets in production
import os
 
def require(name: str) -> str:
    val = os.environ.get(name)
    if not val:
        raise RuntimeError(f"missing required env: {name}")
    return val
 
ENV = os.environ.get("APP_ENV", "development")
MODEL = os.environ.get("AGENT_MODEL", "openai/gpt-4o-mini")  # verify at build
API_KEY = require("OPENROUTER_API_KEY") if ENV == "production" else os.environ.get("OPENROUTER_API_KEY", "")

Rule	Detail
Prefix consistently	`AGENT_*` for app, provider-native names when SDKs expect them
Fail fast in prod	Missing key → non-zero exit before serving traffic
Never log secrets	Redact Authorization headers and connection strings
Document defaults	README or ops runbook lists every var

E - Rotation and Blast Radius

Control	Checklist
Per-environment keys	[ ] Dev / CI / staging / prod isolated
Per-agent keys	[ ] High-risk tools use separate credentials
Spend caps	[ ] Provider and internal budgets set
Rotation drill	[ ] Can replace a key without redeploying code images (config-only roll)
Leak response	[ ] Disable key, rotate, review logs, open incident
Git history	[ ] Secret scanners on CI; rewrite if committed

Pair with Scoping API Keys and Credentials Passed to an Agent.

F - What Must Never Be in the Image or Repo

.env with real keys
Private keys and kubeconfigs
Customer tokens dumped into prompts/logs
"Temporary" prod keys in PR previews
Dockerfile ENV OPENROUTER_API_KEY=sk-...

# .dockerignore / .gitignore essentials
.env
.env.*
!.env.example
*.pem

Ship .env.example with empty or fake values only.

G - Quick Verification Commands

# Fail if a secret pattern is tracked (illustrative)
git grep -nE 'sk-or-|sk-ant-|AKIA' && echo "possible secret" && exit 1 || true
 
# Container should not contain .env
docker run --rm your-image:tag ls -la /app | grep -i env && exit 1 || true
 
# Runtime sees key without printing it
docker run --rm -e OPENROUTER_API_KEY your-image:tag \
  python -c "import os; assert os.getenv('OPENROUTER_API_KEY'); print('key present')"

H - Cross-Target Consistency Checklist

FAQs

Is putting secrets in env vars safe enough?

Env vars are standard but visible to the process and sometimes to platform UIs/logs. Combine with RBAC, short-lived credentials, and never print them. Some teams prefer file mounts with tighter permissions.

Should model names be secrets?

No. Treat model slugs as config. Still pin and change them deliberately for cost/quality.

How do I share config with contractors?

Staging keys only, capped spend, no prod data. Prefer time-limited access.

Can I inject secrets at build time for faster cold starts?

Do not bake secrets into build artifacts. Cold-start gains are not worth permanent key leakage in image layers.

What about client-side keys in web agents?

Browser-exposed keys are public. Use server-side routes or provider proxies; never ship provider master keys to the client.

How often should we rotate?

On leak immediately; otherwise on a schedule your threat model requires (many teams: 90 days or on role change). Automate where the platform allows.

Do orchestrator "sealed secrets" replace discipline?

They help encryption and distribution. You still need naming, least privilege, and rotation process.

Where does this meet production security review?

Use Security Checklist Before Shipping an Agent to Production alongside this page.

Related: Deploying Agents Basics

Related: Containerizing an Agent with Docker

Related: Deploying an Agent to Vercel or AWS Lambda Serverless Functions

Related: Long-Running Agents on a VPS: Process Management and Restarts

Related: Deploying Agents Best Practices

Related: Scoping API Keys and Credentials Passed to an Agent

Related: Security Checklist Before Shipping an Agent to Production

Stack versions: Pins from the category manifest (verify at build): OpenRouter (~315+ models, July 2026 pricing/fees); LangGraph 1.0+; CrewAI 1.14+; Microsoft Agent Framework 1.0; Vercel AI SDK 6; Pydantic AI (latest); LlamaIndex (latest); OpenAI Agents SDK (latest + MCP); MCP (Linux Foundation governance); A2A (HTTP+SSE+JSON-RPC 2.0); Solana @solana/web3.js + @solana/spl-token.

Environment and Secrets Management Across Deployment Targets

Agent deployments fail loudly when keys are missing and fail expensively when keys leak or are shared without caps.

Use this cheatsheet to keep config, secrets, and environments consistent across laptop, serverless, containers, and VPS.

How to Use This Cheatsheet

Treat every row as a gate before promoting a release.
Separate non-secret config (model slug, log level) from secrets (API keys, DB passwords).
Prefer one naming scheme (AGENT_MODEL, OPENROUTER_API_KEY) across all hosts.
Re-audit after adding a new tool that needs credentials.

A - Secret Classes

Class	Examples	Handling
Model provider keys	OpenRouter, OpenAI, Anthropic	Secret store; per-env keys; spend caps
Tool credentials	GitHub PAT, Slack bot, Stripe	Least privilege; scoped to tool identity
Data stores	DB URL, Redis URL	Secret; rotate with app deploy
Signing / session	JWT secret, webhook HMAC	Secret; never log
Non-secret config	`AGENT_MODEL`, `LOG_LEVEL`, `PORT`	Plain env or config map OK

B - Injection by Target

Target	Prefer	Avoid
Local dev	`.env` (gitignored) + direnv/dotenv	Committing `.env`
Docker CLI	`-e`, `--env-file`, orchestrator secrets	`ENV KEY=secret` in Dockerfile
Compose	`${VAR}` from host env or secrets	Hard-coded keys in compose committed files
Vercel	Project env (Production / Preview / Development)	Same key for preview + prod
AWS Lambda	Function env + Secrets Manager/SSM	Long-lived keys when IAM role works
Kubernetes	Secret → env or volume mount	Base64 Secrets without encryption at rest / RBAC
VPS systemd	`EnvironmentFile=` mode `600`	World-readable unit drop-ins with keys
CI	OIDC to cloud / short-lived tokens	Long-lived cloud keys in plain CI vars when avoidable

C - Environment Matrix

Variable	Local	Preview/Staging	Production
`OPENROUTER_API_KEY`	Dev-capped key	Staging key	Prod key
`AGENT_MODEL`	Cheap default	Prod-like	Prod pin
`LOG_LEVEL`	debug	info	info/warn
`ENV` / `APP_ENV`	development	staging	production
Tool write credentials	Mock or sandbox	Sandbox	Prod least privilege

Never let staging share production write credentials "for convenience."

D - Naming and Loading Pattern

# pattern: explicit required secrets in production
import os
 
def require(name: str) -> str:
    val = os.environ.get(name)
    if not val:
        raise RuntimeError(f"missing required env: {name}")
    return val
 
ENV = os.environ.get("APP_ENV", "development")
MODEL = os.environ.get("AGENT_MODEL", "openai/gpt-4o-mini")  # verify at build
API_KEY = require("OPENROUTER_API_KEY") if ENV == "production" else os.environ.get("OPENROUTER_API_KEY", "")

Rule	Detail
Prefix consistently	`AGENT_*` for app, provider-native names when SDKs expect them
Fail fast in prod	Missing key → non-zero exit before serving traffic
Never log secrets	Redact Authorization headers and connection strings
Document defaults	README or ops runbook lists every var

E - Rotation and Blast Radius

Control	Checklist
Per-environment keys	[ ] Dev / CI / staging / prod isolated
Per-agent keys	[ ] High-risk tools use separate credentials
Spend caps	[ ] Provider and internal budgets set
Rotation drill	[ ] Can replace a key without redeploying code images (config-only roll)
Leak response	[ ] Disable key, rotate, review logs, open incident
Git history	[ ] Secret scanners on CI; rewrite if committed

Pair with Scoping API Keys and Credentials Passed to an Agent.

F - What Must Never Be in the Image or Repo

.env with real keys
Private keys and kubeconfigs
Customer tokens dumped into prompts/logs
"Temporary" prod keys in PR previews
Dockerfile ENV OPENROUTER_API_KEY=sk-...

# .dockerignore / .gitignore essentials
.env
.env.*
!.env.example
*.pem

Ship .env.example with empty or fake values only.

G - Quick Verification Commands

# Fail if a secret pattern is tracked (illustrative)
git grep -nE 'sk-or-|sk-ant-|AKIA' && echo "possible secret" && exit 1 || true
 
# Container should not contain .env
docker run --rm your-image:tag ls -la /app | grep -i env && exit 1 || true
 
# Runtime sees key without printing it
docker run --rm -e OPENROUTER_API_KEY your-image:tag \
  python -c "import os; assert os.getenv('OPENROUTER_API_KEY'); print('key present')"

H - Cross-Target Consistency Checklist

FAQs

Is putting secrets in env vars safe enough?

Should model names be secrets?

No. Treat model slugs as config. Still pin and change them deliberately for cost/quality.

How do I share config with contractors?

Staging keys only, capped spend, no prod data. Prefer time-limited access.

Can I inject secrets at build time for faster cold starts?

Do not bake secrets into build artifacts. Cold-start gains are not worth permanent key leakage in image layers.

What about client-side keys in web agents?

Browser-exposed keys are public. Use server-side routes or provider proxies; never ship provider master keys to the client.

How often should we rotate?

On leak immediately; otherwise on a schedule your threat model requires (many teams: 90 days or on role change). Automate where the platform allows.

Do orchestrator "sealed secrets" replace discipline?

They help encryption and distribution. You still need naming, least privilege, and rotation process.

Where does this meet production security review?

Use Security Checklist Before Shipping an Agent to Production alongside this page.

Related: Deploying Agents Basics

Related: Containerizing an Agent with Docker

Related: Deploying an Agent to Vercel or AWS Lambda Serverless Functions

Related: Long-Running Agents on a VPS: Process Management and Restarts

Related: Deploying Agents Best Practices

Related: Scoping API Keys and Credentials Passed to an Agent

Related: Security Checklist Before Shipping an Agent to Production

Stack versions: Pins from the category manifest (verify at build): OpenRouter (~315+ models, July 2026 pricing/fees); LangGraph 1.0+; CrewAI 1.14+; Microsoft Agent Framework 1.0; Vercel AI SDK 6; Pydantic AI (latest); LlamaIndex (latest); OpenAI Agents SDK (latest + MCP); MCP (Linux Foundation governance); A2A (HTTP+SSE+JSON-RPC 2.0); Solana @solana/web3.js + @solana/spl-token.

Environment and Secrets Management Across Deployment Targets

Buscar en la documentación

Environment and Secrets Management Across Deployment Targets