Launchdarkly

Documentation at docs.launchdarkly.com is comprehensive and well-organized by concept (flags, targeting, experimentation) and SDK (server-side, client-side, edge). The REST API reference is auto-generated from OpenAPI with try-it functionality. SDK documentation is thorough — each SDK has its own getting-started guide, configuration reference, and migration guide for version upgrades. The Learning Center provides conceptual guides on feature management practices. Community resources include a Slack community and conference talks. The documentation's strength: it bridges the gap between 'how the API works' and 'how to practice feature management' — useful for agents making flag management decisions. SDKs cover every major language and platform. The API explorer enables interactive testing.

LaunchDarkly is the dominant feature management platform, and its value for agents lies in two dimensions: flag evaluation (should this feature be on/off for this user?) and flag management (create, modify, and schedule feature flags programmatically). The SDK-based evaluation model means agents using LaunchDarkly for flag checks get local evaluation with streaming updates — no API call per flag check, sub-millisecond performance. The REST API covers flag creation, targeting rule management, environment configuration, segment management, and audit log access. For agents orchestrating feature releases — progressive rollouts, canary deployments, kill switches — LaunchDarkly provides programmatic control over the entire lifecycle. Experimentation features enable A/B testing with statistical analysis. The main consideration: LaunchDarkly is enterprise-priced, and the SDK model requires initialization context that adds complexity versus simple API-based flag services.

REST API uses personal or service access tokens with role-based scoping: reader, writer, admin, or custom roles. Custom roles enable fine-grained permissions: flag-level, project-level, or environment-level access control. SDK keys are separate: SDK keys for server-side evaluation, mobile keys for mobile SDKs, and client-side IDs for browser SDKs. SDK keys provide read-only flag configuration access. The access token model supports the principle of least privilege well — agents can have tokens scoped to specific projects or actions. Service tokens are recommended for automated/agent access. Relay proxy enables air-gapped SDK evaluation for high-security environments. OAuth is available for marketplace integrations.

Two integration surfaces: SDKs for flag evaluation and REST API for flag management. Server-side SDKs (Go, Node, Python, Java, Ruby, .NET, etc.) stream flag configurations and evaluate locally — no network call per flag check. This is the fastest evaluation model among feature flag platforms. REST API at app.launchdarkly.com/api/v2/ covers projects, environments, feature flags, segments, users, and audit logs. Flag creation supports boolean, multivariate string, number, and JSON variation types. Targeting rules use user attributes with AND/OR logic. Scheduled flag changes enable future-dated rollouts. The API uses semantic patching for flag updates — PATCH requests use JSON Patch format with instructions array, which is powerful but non-obvious. Webhooks notify on flag changes. Flag status endpoints track stale and inactive flags.

REST API errors return JSON with code and message. Standard HTTP status codes. Rate limits: 10 requests per second per access token for writes, higher for reads. The SDK evaluation model is inherently resilient: SDKs cache flag configurations locally and continue evaluating even if the streaming connection drops. Default values are returned for unknown flags. Flag prerequisites prevent dependent flags from activating out of order. The audit log tracks every flag change with actor and timestamp. Flag status tracking identifies stale flags (never evaluated) for cleanup. The main risk for agents: the semantic patching API means malformed PATCH instructions silently fail to apply changes — agents must validate patch operations carefully. Environment-specific flag states prevent accidental production changes from staging.