toggleflag

Lightweight feature flags for Python — no server, no dependencies.

import toggle

flags = toggle.Flags()

flags.define("new_checkout", default=True)
flags.define("experiment_v2", variants=["control", "treatment"], rollout_pct=30)

flags.is_on("new_checkout")                              # True
flags.variant("experiment_v2", user_id="u1")            # "treatment" (30% of the time)

with flags.context(user_id="u1", country="US"):
    flags.is_on("us_only_feature")                       # checks per-user + per-country

---

Why toggleflag?

Feature flags are essential for modern software delivery. They enable teams to ship code behind toggles, decouple deployment from release, run A/B experiments, and instantly disable broken features in production.

Existing solutions have significant trade-offs:

Concern	toggleflag	LaunchDarkly	Unleash	django-flags
Requires a server	❌ No	✅ Yes	✅ Yes	❌ No
External dependencies	Zero	SDK dependency	SDK dependency	Django framework
Setup time	< 1 minute	Account + SDK setup	Self-hosted cluster	Django project
Per-user targeting	✅	✅	✅	Limited
Percentage rollout	✅ Deterministic	✅	✅	❌
Country targeting	✅	✅ (paid)	✅ (enterprise)	❌
A/B testing variants	✅	✅ (paid)	✅	❌
Context propagation	✅	✅	✅	❌
Live config reload	✅ File watching	✅ Streaming	✅ Polling	✅
Thread-safe	✅	✅	✅	✅
Pricing	Free (MIT)	$10–$500+/mo	Free (OSS) / Enterprise	Free
Python-only	✅	Multi-language	Multi-language	Python (Django)

toggleflag fills a gap: a zero-dependency, pure-Python feature flag library that supports sophisticated targeting rules without requiring any infrastructure.

---

Installation

pip install toggleflag

That's it. No servers to configure. No databases to provision. No accounts to create.

---

Quick Start

Boolean Flags

import toggle

flags = toggle.Flags()

flags.define("dark_mode", default=True)
flags.define("maintenance_mode", default=False)

flags.is_on("dark_mode")          # True
flags.is_off("maintenance_mode")  # True

Percentage Rollout

flags.define("new_checkout", default=False, rollout_pct=50)

# Deterministic: same user always gets the same result
flags.is_on("new_checkout", user_id="alice")   # True (consistent)
flags.is_on("new_checkout", user_id="bob")     # False (consistent)

A/B Testing with Variants

flags.define("pricing_page", variants=["control", "variant_a", "variant_b"], rollout_pct=40)

# Returns a variant string, or None if the user is outside the rollout
flags.variant("pricing_page", user_id="u1")  # "variant_a"
flags.variant("pricing_page", user_id="u2")  # None (outside 40% rollout)

Per-User Targeting

# Beta access for specific users
flags.define("beta_features", default=False, user_ids=["alice", "bob", "charlie"])

flags.is_on("beta_features", user_id="alice")   # True
flags.is_on("beta_features", user_id="dave")    # False

# Per-user overrides (force on/off regardless of other rules)
flags.define("experiment", default=False, rollout_pct=10,
             overrides={"alice": True, "bob": False})

flags.is_on("experiment", user_id="alice")  # True (forced on)
flags.is_on("experiment", user_id="bob")    # False (forced off)

Country Targeting

flags.define("gdpr_consent_flow", default=True, countries=["DE", "FR", "GB"])

flags.is_on("gdpr_consent_flow", country="US")  # False
flags.is_on("gdpr_consent_flow", country="DE")  # True

Context Propagation

with flags.context(user_id="alice", country="US"):
    # All flag checks within this block use alice/US context
    flags.is_on("beta_features")      # True (alice is in user_ids)
    flags.is_on("us_only_promo")      # True (US country match)
    flags.variant("experiment_v2")    # Deterministic variant for alice

    # Nested contexts merge with parent
    with flags.context(country="DE"):
        flags.is_on("gdpr_consent_flow")  # True (DE, alice's user_id preserved)

# Explicit arguments override context
with flags.context(user_id="alice"):
    flags.is_on("beta_features", user_id="bob")  # False (explicit arg wins)

File-Based Configuration

{
  "new_checkout": {"default": true, "rollout_pct": 50},
  "dark_mode": {"default": true},
  "experiment_v2": {
    "variants": ["control", "treatment"],
    "rollout_pct": 30,
    "countries": ["US", "GB"]
  },
  "beta_features": {
    "default": false,
    "user_ids": ["alice", "bob"]
  }
}

flags = toggle.Flags()
flags.load_file("flags.json")

Environment Variables

export TOGGLE_DARK_MODE=true
export TOGGLE_NEW_CHECKOUT=50
export TOGGLE_EXPERIMENT=50:treatment,control

flags = toggle.Flags()
flags.load_env()

# Custom prefix
flags.load_env(prefix="APP_")  # reads APP_DARK_MODE=true, etc.

Auto-Reload File Changes

watcher = flags.watch("flags.json", poll_interval=2.0)

# In your request handler or background loop:
if watcher.changed():
    flags.load(watcher.read())
    print("Flags reloaded!")

---

Deterministic Rollout

Percentage-based targeting in toggleflag is deterministic — the same user always gets the same result, across requests, servers, and restarts.

This works by computing MD5(flag_name + ":" + user_id) % 100 and comparing against rollout_pct. The result depends only on the flag name and user identifier, not on random state.

flags.define("experiment", variants=["control", "treatment"], rollout_pct=30)

# These calls always return the same value for the same user
flags.variant("experiment", user_id="alice")  # Always "treatment"
flags.variant("experiment", user_id="bob")    # Always None

Why MD5? MD5 is fast (important for request-time evaluation), uniformly distributed, and deterministic. While not cryptographically secure, that's irrelevant here — we're bucketing users, not protecting secrets.

No user ID? When no user identifier is provided, the system uses a default sentinel value, meaning all anonymous users get the same result. For meaningful rollouts, always provide a user_id.

---

API Reference

Flags()

Create a new flag engine instance. Thread-safe.

flags.define(name, *, default=False, rollout_pct=?, variants=None, user_ids=None, countries=None, overrides=None)

Define a feature flag.

Parameter	Type	Default	Description
name	str	—	Flag identifier
default	bool	False	Default on/off state (used when no rollout_pct is set)
rollout_pct	int	auto	0–100+ percentage of users who see the flag
variants	list[str]	[]	Variant names for A/B testing
user_ids	list[str]	None	Allow-list of user IDs
countries	list[str]	None	Allow-list of country codes
overrides	dict	{}	Per-user forced values {"user_id": True/False/"variant"}

flags.is_on(name, *, user_id=None, country=None) → bool

Check if a flag is on for the given context.

flags.is_off(name, **kwargs) → bool

Inverse of is_on.

flags.variant(name, *, user_id=None, country=None) → str | None

Return the A/B variant for the user, or None if outside rollout.

flags.context(*, user_id=None, country=None, **extra) → ContextManager

Push a temporary evaluation context. Nestable — child contexts merge with parents.

flags.load(data: dict) / flags.load_file(path) / flags.load_env(prefix="TOGGLE_")

Bulk-load flag definitions from a dict, JSON/YAML file, or environment variables.

flags.watch(path, poll_interval=2.0) → FileWatcher

Create a file watcher for live config reloading.

flags.list_flags() → list[str] / flags.get(name) → FlagDef | None / flags.undefine(name) / flags.clear()

Introspection and management methods.

---

Design Philosophy

1. Zero dependencies. Pure Python 3.10+. No external packages. No database. No network calls. pip install and go.

2. Deterministic by default. Same user, same flag, same result — always. No random state to corrupt or drift.

3. Configuration flexibility. Programmatic API, JSON/YAML files, environment variables, or any combination. Use what fits your deployment.

4. Targeting composure. User IDs, countries, percentages, and overrides compose naturally. Each rule is a filter; they stack.

5. No infrastructure required. Unlike LaunchDarkly or Unleash, toggleflag needs no server, database, or cluster. A JSON file is sufficient for production use.

6. Thread-safe. All flag reads and writes are protected by locks. Safe for multi-threaded web servers (Django, Flask, FastAPI).

7. Explicit is better than implicit. Flags must be defined before use. Undefined flags return False rather than raising errors — fail soft in production.

---

Use Cases

A/B Testing

flags.define("pricing_page", variants=["control", "variant_a", "variant_b"], rollout_pct=50)

variant = flags.variant("pricing_page", user_id=current_user.id)
if variant == "variant_a":
    return render_pricing_v2(request)
return render_pricing(request)

Phased Rollout

# Week 1: 10% of users
flags.define("new_api", default=False, rollout_pct=10)

# Week 2: Increase to 50%
flags.define("new_api", default=False, rollout_pct=50)

# Week 3: Full launch
flags.define("new_api", default=True)

Canary Releases

flags.define("v2_engine", default=False, user_ids=["alice", "bob", "qa_team"])

if flags.is_on("v2_engine", user_id=request.user.id):
    return v2_handler(request)
return v1_handler(request)

Kill Switches

# Disable a feature instantly without redeploying
flags.define("external_payments", default=True)

# In a config file — flip to false to disable
# {"external_payments": {"default": false}}

Country-Specific Features

flags.define("eu_compliance", default=True, countries=["DE", "FR", "IT", "ES", "GB"])
flags.define("us_tax_form", default=True, countries=["US"])

Beta Programs

flags.define("beta_ai_features", default=False, user_ids=beta_users)

---

Testing

pip install -e ".[dev]"
pytest tests/ -v

87 tests covering all targeting rules, context propagation, file/env loading, thread safety, and edge cases.

---

License

MIT © Ravi Teja Prabhala Venkata