tashfeenahmed/freellmapi v0.7.0

on GitHub

Highlights

The largest release since 0.6.0: 39 commits covering routing, the OpenAI-compatible surfaces, key management, analytics, security hardening, and the desktop app.

Routing and reliability

• Per-model routing weight overrides via the MODEL_ROUTING_OVERRIDES env var: demote, promote, or exclude individual models from automatic routing without disabling them (#747, #738)
• Failed providers are skipped as a whole on 5xx/timeout/transport errors instead of one key at a time (#817)
• Models that repeatedly fail are auto-cooled-down, extending the penalty beyond 429s (#806)
• Provider back-off stated in the error body is honored, not just the Retry-After header (#798)
• The exploration probe only considers models that can actually serve the request, and respects weight-0 overrides (#847)
• Opt-in X-Fallback-Detail response header with per-hop failover timings (#792)

OpenAI-compatible API

• /v1/responses aligned with the official Responses API spec: unique streaming output_index, all official input item types accepted, assistant-turn merging, system-message hoisting, and a clear 422 for computer use (#844, #842, #850)
• Streaming responses now record the provider's real token usage instead of a character-count estimate, making analytics and rate-limit accounting accurate (#845)
• Opt-in schema validation of tool-call arguments across all surfaces, failing over to the next model on invalid arguments (#802, #821)
• Double-encoded tool arguments below the top level are repaired (#794)
• Reasoning tokens are counted and TTFB is recorded on the first reasoning token (#778, #764)
• Session reasoning_content dropped by clients on replay is restored (#805)
• Anthropic surface no longer drops document content blocks (#793)

Keys and providers

• Per-key proxy override: each API key can route through its own exit (#819, #590)
• Bulk enable/disable/delete for keys within a group (#816)
• Providers that need a key but have none added are flagged in the dashboard (#777)
• AnyAPI added as an OpenAI-compatible platform (#772)
• Full-key copy on the desktop build without password re-verification (#810)

Dashboard and analytics

• Automatic update check with a release-notes dialog (#782)
• Timeline charts bucket and render in the viewer's local timezone (#762, #804)
• Custom-provider recent calls show the user's key label (#815)
• Unified output-token cap setting rescues excessive client max_tokens (#783)
• zh-CN translation polish and full-width dash sweep in zh-CN/zh-TW (#765, #748)

Security hardening

• Database directory, backup blobs, and restored databases are owner-only, with real permissions on Windows and WAL sidecars covered (#791, #795, #796)

Desktop

• Electron 38.8.6 for macOS 26 Tahoe window and tray rendering (#808)
• Windows zip shipped alongside the installer (#773)
• LF line endings pinned, fixing a CSP hash mismatch on Windows (#784)

macOS note: the DMG is unsigned; right-click the app and choose Open on first launch.

Full Changelog: v0.6.9...v0.7.0