v0.3.0

This release introduces Interactive Web Bench Packs, a major expansion of BenchLocal beyond table-based benchmark runs.

Highlights

• Added support for hosted type: "web" Bench Packs.
• Added the new @benchlocal/web-sdk package for web benchmark apps.
• Web Bench Packs can run inside BenchLocal while using the user’s local saved providers, models, credentials, sampling settings, and inference flow.
• Provider credentials stay inside BenchLocal. Hosted benchmark pages act as the presentation and interaction layer.
• Added local bridge APIs for model selection, inference, streaming, run state, history payloads, and artifacts.
• Added start/stop state handling so BenchLocal can reflect when a web benchmark is actively running.
• Added BenchLocal-side stop support for web benchmarks.
• Added official documentation: docs/interactive-web-benchpacks.md.

Web Bench Packs

• BenchLocal can now install web-hosted packs from manifests.
• Web pack manifests support version gating through requirements.benchlocal.minVersion.
• Registry installs now support hosted web sources.

Developer SDK

• Added @benchlocal/web-sdk@0.3.0.
• Provides the browser-side API for detecting BenchLocal, listing selected models, running inference, streaming responses, saving history, and reacting to BenchLocal start/stop controls.

v0.2.9

This release adds visual polish and theme improvements.

New

• Added a new built-in Black theme.

Improved

• Polished the share result card score panel.
• Enlarged and right-aligned the score number.

v0.2.8

This release adds shareable benchmark result cards.

New

• Added a Share action for model benchmark results.
• Added a share card preview modal.
• Supports saving the result card as a high-resolution PNG.
• Supports copying the result card image to the clipboard when the platform allows it.

v0.2.7

This release adds BenchLocal’s first agent-facing control surfaces, including a local HTTP/SSE Agent API and a standard MCP endpoint, so AI agents can control benchmark workflows while the desktop UI updates in real time.

Highlights

• Added local Agent API for automation and AI agents.
• Added MCP Streamable HTTP endpoint with benchlocal_* tools and state resources.
• Added real-time SSE event stream for run progress and state changes.
• Added Agent Access settings UI with token, guide URL, OpenAPI URL, and MCP URL.
• Added detailed developer reference: docs/agent-control-api.md.
• Added README guidance for using Agent Access, HTTP/SSE, and MCP.

UI Improvements

• Replaced top notification banners with bottom-left auto-dismissing toasts.
• Deduped repeated stop-run notifications.
• Added checkbox selection to test history rows.
• Added header select-all for histories.
• Added Delete selected for selected histories.
• Added a new theme: Sunforge.

Developer Notes

• Agent API and MCP now share the same controller path as the desktop UI.
• Provider secrets are redacted from API/MCP responses.
• Long-running agent/MCP run commands return quickly while progress remains visible in the UI.
• The new API/MCP reference includes an extension checklist so future UI features can be exposed consistently.

v0.2.6

Added

• Added model/provider availability handling so unavailable local models can be skipped temporarily and completed later with resume.
• Added online/offline model status indicators in test result tables.
• Added retry controls in history mode for provider errors and failed results.
• Added Duplicate actions for Providers and Models. Duplicating a provider creates only the provider record, not linked models.
• Added double-click support in the New Tab Bench Pack picker to open a selected Bench Pack.

Improved

• Provider names are now shown with friendly display names instead of internal UUID-style IDs in model and error displays.
• Empty/offline model score cells now show — instead of 0 when no result data exists.
• Settings screen spacing was refined so content scrolls cleanly without edge gaps.

Fixed

• Provider error classification now trusts HTTP response status codes instead of scanning response text.
• Provider request timeout handling was fixed.
• Fixed development-mode Electron shutdown so npm run dev can quit cleanly.

v0.2.5

• Added repeated test runs with majority voting, supporting 1, 3, 5, 7, and 9 runs per test. Fixes #10.
• Added provider/network error classification with yellow status styling, plus retry actions for provider errors and failed results. Fixes #7.
• Added scenario result timings in the table for new runs, while keeping older results clean when timing data is unavailable. Fixes #11.
• Added presence_penalty sampling override support. Fixes #9.
• Clarified sampling defaults so blank sampling fields are omitted and backend/provider defaults are used. Fixes #12.
• Restored tab right-click duplicate behavior.
• Added a fresh-state action for resumable runs, keeping selected models while clearing partial run state.
• Result summaries now show provider display names instead of internal provider UUIDs.
• Fixed macOS Cmd+Q so it quits the app directly instead of only closing the window first.

v0.2.4

BenchLocal v0.2.4

Changes

• Added Hugging Face provider support.
• Fixed BenchLocal inference relay JSON response headers so Docker verifiers no longer fail with terminated when providers return compressed JSON responses that the relay buffers and rewrites.

v0.2.3

Highlights

• Added model registry filters in Settings -> Models for provider, group, and text search.
• Fixed model alias editing so spaces can be typed normally.
• Stabilized result-table column highlighting during parallel runs.
• Fixed OrbStack Docker detection for newer OrbStack installs.
• Fixed Bench Pack installation on Linux setups where temp storage and home storage are on different filesystems.

Notes

• This release is intended to verify the new GitHub Releases self-update flow from v0.2.2 to v0.2.3.
• Users already on v0.2.2 should be able to update through BenchLocal's in-app updater.
• Users on v0.2.1 or older still need to install manually first.

Verification

• Built macOS, Windows, and Linux release artifacts.
• macOS release build is signed and notarized.
• Verified updater metadata files are generated for all platforms.

v0.2.2

Highlights

• Added in-app self-update support.
• Added update status and restart controls in the BenchLocal UI.
• Improved verifier preparation controls, including cancellation and Docker image cleanup handling.
• Updated macOS behavior so BenchLocal quits when the main window is closed.

Notes

• This is the first BenchLocal release that includes self-update support.
• Users on v0.2.1 still need to install v0.2.2 manually.
• After installing v0.2.2, future releases can be detected and installed from inside the app.

Verification

• macOS release build is signed and notarized.
• Built macOS, Windows, and Linux release artifacts.
• Verified updater metadata files are generated for GitHub Releases.

v0.2.1

BenchLocal 0.2.1

BenchLocal 0.2.1 is a client-focused update that improves long-running Bench Pack workflows, history handling, replay, and settings usability.

Highlights

• Resume incomplete Bench Pack runs from history without creating a new run entry
• Restore loaded results when reopening the app, including resumable incomplete runs
• Preserve partial progress when quitting during an active run
• Add replay mode for completed runs with Shift + Open from Test Histories
• Auto-follow the currently running scenario while still allowing manual column pinning
• Add per-scenario retry polish so retries behave like live activity and refresh scores immediately
• Improve settings tables on smaller windows:
  • sticky action column
  • visible horizontal scrollbar
• Persist the BenchLocal window size and reopen at the last used size
• Improve small-screen desktop behavior

Notes

• This is a BenchLocal client release.
• @benchlocal/core and @benchlocal/sdk remain at 0.2.0 in this release.

Downloads

• macOS Apple Silicon: DMG and ZIP
• Windows x64: EXE and ZIP
• Linux x64: AppImage and tar.gz