The biggest release yet — a month of work spanning 9 new free providers, an Anthropic-compatible API, generative media, an MCP server, structured outputs, and a large reliability + security hardening pass.
New providers (9)
SEA-LION, NavyAI, NaraRouter, Aion, Requesty, AI Horde, Routeway, BazaarLink, AINative — all with genuine free tiers, routed behind the same single key.
New APIs & surfaces
- Anthropic-compatible API —
/v1/messagesworks with Anthropic SDKs and tools (Claude Code included), with system-role fixes and Gemini thought-signature preservation (#361, #362, #493) - Generative media — image generation + audio/TTS endpoints, with Image and Audio tabs in the model browser (#358, #359)
- MCP server at
/mcp— gateway introspection for coding agents (#517) - Structured outputs —
json_schemaforwarded to providers, with JSON healing, routing filter, and format-ignore failover when a provider can't comply (#514, #516, #520) - Legacy completions endpoint (#340) and embeddings
dimensionsparameter for MRL truncation (#393) - Self-hosted OpenAPI reference at
/v1/docs+/v1/openapi.json(#482) - Opt-in response cache for
/v1/chat/completions(#481)
Routing & reliability
- Unified the four provider fallback loops; 401 key rotation, retry budgets, truncation policy, rich exhaustion diagnostics (#483, #484)
- Per-request token budget + failover circuit breaker (#511)
- Client-disconnect abort +
X-Fallback-Trailheader on every surface (#518) - Post-sleep wake recovery: drift detection, socket flush, immediate key re-probe (#515)
- Multi-key improvements: pooled monthly token budgets scale with key count, embeddings/media load balancing, quota fixes (#457, #459, #479)
- Provider timeout tuning from a live all-provider sweep; 60s chat timeout for reasoning-model TTFB (#519, #521)
- Fusion mode: tool-call support and model-family panel diversity (#331)
- Router penalty inspector + quota observability (#344)
Security
- First-run setup code; unified key moved out of the DB into a key file; CSV export hardening (#477)
- Custom-provider SSRF guard, plus two follow-up bypass closures (#443, #450)
Dashboard & desktop
- Keys page redesign with tabs, provider key import/export, custom models managed from the Keys page (#327, #467, #473)
- Cmd/Ctrl+K command palette, models nav dropdown, first-run checklist, skeletons, error boundary + toasts, 404 page
- Consolidated provider detail pages with API usage for image/audio/embeddings (#360)
- Playground: system prompt support + real error bubbles (#390)
- Desktop: LAN access toggle (bind 0.0.0.0), app version in popover (#421, #447)
- i18n: Italian added; complete coverage across all six locales (#367)
Under the hood
- File-per-migration DB system,
Db/DbStatementinterfaces,Schedulerabstraction with lifecycle tests (#265, #397) - Runtime-capability guards + centralized Config (#375); env-drift warning at startup (#494)
- Catalog controls, persistence backup, and declarative config
Full diff: v0.4.1...v0.5.0
Thanks to community contributors: @chongjiazhen, @LoneRifle, @vjsai, @Tazrif-Raim, @kairwang01 and everyone who reported issues.
macOS note: the DMG is not yet notarized — on first launch use right-click → Open (or
xattr -d com.apple.quarantine /Applications/FreeLLMAPI.app).