●Product

Three primitives. One slice axis per layer.

Customer → Product → Tenant. The Customer is your workspace — billing, users, allowed domains. The Product is what your SDK authenticates as — its own apiKey, its own flag namespace, its own audit stream. The Tenant is the slice your flag value resolves for — a B2B customer organisation, a brand variant of one codebase, a region, a cohort. Same data shape, same override semantics, same audit boundary regardless of what you're slicing by. Every flag, every rule, every audit row is scoped at the storage layer — bypass is impossible because the row can't be written without all three.

●Architecture

The whole platform is built around three data layers.

MVP runs on Firebase. A custom data plane is a post-PMF decision, not day one.

Layer 04SDK

evaluate() subscribe() audit()

Layer 03Rules

flag tenant override rollout approval

Layer 02Tenant

tenant_id env scope region

Layer 01Storage

firestore onSnapshot · change log cloud functions

●SDK

Per-Product apiKey. Per-tenant evaluation.

Each Product has its own apiKey — the SDK uses it to identify which Product's flags to fetch. For multi-tenant Products, the evaluation API also requires a tenantId, enforced server-side. iOS, Android, and Web SDKs ship today, pushing flag changes in real time via Firestore listeners (~1s, no polling). Web ships as two packages — @argus/client for the browser (real-time) and @argus/node for server-side pull / SSR. Edge resolution lands in v1.5.

typescript · argus-sdk@v0

// scoped, deterministic, reversible
const flag = await argus.evaluate({
  key:     "new_billing_v2",
  tenant:  ctx.tenantId,      // required
  env:     "prod",
});

if (flag.on) renderNewBilling();

// → audit row written, scoped to tenant.
// → reversible from the dashboard in one click.
// → impossible to evaluate without a tenant.

●Web SDK

Drop Argus into the browser and the server.

The Web SDK ships as two packages that share one read-scoped apiKey. @argus/client runs in the browser: it seeds flag values with a single resolveFlags pull, then subscribes to a Firestore listener so changes land in real time (~1s, no polling). @argus/node runs on the server for SSR and request-scoped evaluation, and serialises resolved flags so the browser hydrates without a second round trip. Same shape as the iOS and Android SDKs — one apiKey per Product, an optional tenantId for multi-tenant resolution.

typescript · @argus/client

// browser · real-time via Firestore listener
import { init } from "@argus/client";

const argus = await init({
  apiKey:   import.meta.env.ARGUS_KEY,
  baseUrl:  "https://api.your-argus.cloud",
  tenantId: "acme_retail",   // optional
});

if (argus.getFlag("new_billing_v2", false))
  renderNewBilling();

// re-renders when a flag flips, live
argus.onChange((flags) => paint(flags));

// React: import { useFlag } from "@argus/client/react"
argus.close(); // detach the listener

typescript · @argus/node

// server · request-scoped pull + SSR hydration
import { createClient } from "@argus/node";

const argus = createClient({
  apiKey:  process.env.ARGUS_KEY,   // same key
  baseUrl: "https://api.your-argus.cloud",
});

const ctx = { userId: user.id, tenantId: "acme_retail" };
const on = await argus.getFlag("new_billing_v2", false, ctx);

// hand the resolved set to the browser, no re-fetch
res.locals.argusState = argus.serialize(ctx);
// → @argus/client picks it up to hydrate instantly.

baseUrl is your Cloud Functions origin. One read-scoped apiKey covers both packages — there is no public / secret split. The init and resolution API shown here is shipped today; the npm packages publish with v1 — until then a preview build is available from the SDK repository.

●MCP

Manage flags from your AI coding agent.

Argus ships a local MCP server (argus-mcp) so an AI coding agent — Claude Code, Cursor, and the like — can manage feature flags from inside a coding session: list, create, and publish flags, set per-environment values, and toggle them, without leaving the editor.

01 · Governed by default

New flags land as DRAFT — invisible to SDKs until published.

Changing a prod value requires explicit confirmation. The agent gets reach, not a blank cheque.

02 · Every mutation audited

The agent writes to the same audit log as a human.

Each change records actor, timestamp, env, tenant, flag, before, and after — tagged as the agent. Append-only, same as any other change.

03 · Founder / local today

Runs as a local stdio server, authenticated with a local service-account key.

No per-agent tokens yet. A productized, multi-tenant MCP with per-agent revocable tokens is on the roadmap.

●Tenant overrides

Resolution order, plainly stated.

Tenant override documents are first-class — not a special case of segments. They have their own audit stream and resolution lane.

→

Default

flag default value at evaluation time

→

Flag rule

global rule (env, segment, version target)

→

Tenant override

tenant-owned override document · audited

→

Rollout

per-tenant % · deterministic per-user bucket (seed + userId)

→

Resolved value

flag.on · with source metadata in response

→

Audit row

append-only · tenant-scoped · reconciled

●Approvals & RBAC

Two engineers approve prod. Roles scope to tenant.

01 · Two-engineer approval

Production-impacting changes require a second reviewer.

Approval is a flow, not a checkbox. Requested → assigned → approved → applied. Each step writes an audit row.

approval · requested · by shahin just now

awaiting 1 of 2 · reviewers marco, roberto —

02 · Tenant-scoped RBAC

Three roles, each a strict superset of the last.

A contributor changes dev and proposes everything higher. An approver approves changes and sets dev / staging directly. A super_user owns prod and user management.

Role	dev	staging	prod	approve	users
contributor	●	○	○	○	○
approver	●	●	○	●	○
super_user	●	●	●	●	●

●Audit log

Append-only. Product- and tenant-scoped. Source metadata travels.

Every resolution response carries the source of the value (default · rule · override · rollout). Every change writes a row you can query by product and tenant.

argus / audit · tenant=98chimp streaming

flag=new_billing_v2 tenant=98chimp env=prod before=off after=50% actor=shahin just now

approval=applied flag=new_billing_v2 approvers=marco, shahin 2m

override=created flag=kyc_lite tenant=98chimp value=on 14m

rbac=grant role=approver scope=tenant:98chimp subject=roberto@98chimp.com 38m

4 rows · last 1h · exportable to your siem append-only

●Capability matrix

Multi-tenancy as the first row, not a checkbox.

Honest comparison against the alternatives most teams already run. We update this page as the tools ship changes.

Capability	Argus	LaunchDarkly	Statsig	Flagsmith	DIY
Tenant as first-class primitive	●	○	◐	◐	○
Scoped audit log per tenant	●	◐	◐	○	○
Two-engineer prod approval	●	✓	○	○	○
Deterministic per-tenant rollout (hash)	●	✓	✓	◐	◐
Self-host parity (control + data plane)	●	○	○	✓	✓
Pricing not coupled to end-customer MAU	●	○	○	◐	✓
● built ◐ partial ✓ via add-on ○ none · source · manus apr 2026, vendor docs apr 2026

●Roadmap

Plain text. Honest horizons.

Shipping now

v0 · MVP on Firebase

Tenant-scoped evaluation API
Per-tenant override documents
iOS, Android + Web SDKs (real-time push)
Two-engineer prod approval
Tenant-scoped RBAC
Immutable audit rows
Reversible tenant-scoped rollback

Next 60 days

v1 · with design partners

Tenant feature matrix export (CSV + API)
Audit feed export to customer SIEM
Self-host helm chart (control + data parity)

On the roadmap

v1.5+ · with proof

Edge resolution
Code-reference scanning · dead flag detection
SOC 2 audit export packaging
Operational AI · tenant rollout suggestions from history

●What's next

Start free, or see the pricing.

Free is $0 with no card. Growth is $99/month, with a per-tenant discount curve for multi-tenant portfolios. The full breakdown is on the pricing page.

Start free → See pricing →