Atlas Vision MCP v1.4.0

Highlights

Atlas Vision MCP v1.4.0 adds secure mid-turn image interception for Pi text-only models. When a Pi tool such as read returns a screenshot, Atlas can analyze it and append structured evidence before the model's next turn—without requiring a separate manual vision call.

Pi tool-result interception

• Uses explicit tool-result image blocks as the canonical image source.
• Analyzes each unique image once, avoiding duplicate provider calls when read also supplies an input path.
• Falls back to a path only for successful read results with no image block.
• Never scans arbitrary ls, find, shell, or text output for image paths.
• Preserves ATLAS_ALLOWED_DIRS and does not auto-expand the upload boundary.
• Uses isolated temporary directories with cleanup on success, failure, cancellation, and parallel execution.
• Propagates Pi cancellation to the provider request boundary.

Gallery preview

The Pi package-gallery artwork has been redesigned with clearer hierarchy, stronger small-size legibility, and fixed layout zones that prevent overlapping content. The editable SVG source is included alongside the PNG.

Verification

• Release commit CI: build, lint, typecheck, and the complete Linux unit suite passed.
• Official Pi createReadTool() integration verifies that one returned image produces one Atlas call.
• Package dry run verified the expected README.md, dist/, extensions/, and hooks/ contents.
• Local provider doctor reported healthy connectivity.

Upgrade

npm install -g atlas-vision-mcp@1.4.0

Pi users can install or update with:

pi install npm:atlas-vision-mcp

Contributors

Thank you to @jagaliano for contributing the original Pi tool_result interception implementation and identifying the correct extension point. The contribution and maintainer hardening are preserved in the merge history for #6.

Full changelog: v1.3.0...v1.4.0

v1.3.0

Highlights

• Allow http:// provider base URLs for loopback and private-network endpoints, while continuing to require HTTPS for public hosts.
• Resolve ATLAS_ALLOWED_DIRS roots through symlinks before path checks, fixing valid image paths on platforms such as macOS where /tmp and /var may resolve elsewhere.

Validation

• GitHub CI: all required jobs passed (build, lint, typecheck, tests; live-provider jobs skipped without credentials).
• Release publish checks: 51 test files / 470 tests passed, typecheck and build passed.
• Published to npm as atlas-vision-mcp@1.3.0 with the latest dist-tag.

Full details: CHANGELOG.md.

v1.2.0

Added

• Interactive Pi users can now run /atlas off, /atlas on, /atlas auto, or /atlas status to control Atlas image interception in the current session.
• auto preserves existing capability-based routing; on and off override it without editing environment files.

Validation

• GitHub CI: passed
• Publish workflow: build, lint, typecheck, test, and npm publish all passed
• npm: atlas-vision-mcp@1.2.0

atlas-vision-mcp v1.1.0

Atlas Vision MCP v1.1.0

Atlas Vision MCP 1.1.0 expands the Responses API integration surface, improves long-running tool reliability, and gives callers finer control over high-effort image analysis without changing the default lightweight workflow.

Highlights

• Added Volcengine Ark compatibility for the openai-responses provider, including provider-specific image detail handling and configurable thinking / store request fields.
• Added per-call reasoning_effort and model overrides to analyze_image, allowing a caller to escalate quality on the same image without reconfiguring the server.
• Added progress-heartbeat protection for long-running MCP tool calls so compliant clients can keep requests alive beyond the SDK's default 60 second timeout.
• Added a prebuild safety guard that blocks rebuilding while a live Atlas MCP server is still running from dist/.

Configuration and runtime changes

• Added VISION_RESPONSES_THINKING, VISION_RESPONSES_EFFORT, and VISION_RESPONSES_STORE.
• Raised the default VISION_TIMEOUT_MS from 60s to 300s for reasoning-enabled vision calls.
• Updated cache key scoping so escalated retries do not reuse lower-effort or different-model cached responses.

Correctness fixes included in the merged follow-ups

• Scoped Volcengine xhigh detail handling to the Responses provider path instead of changing the shared provider mapping globally.
• Propagated per-call model overrides across Gemini, OpenAI-compatible, Claude, and Responses providers so the MCP tool contract matches runtime behavior.
• Landed the formatting and import-order fixes required to restore CI cleanliness before release.

Validation

• pnpm build
• pnpm lint
• pnpm typecheck
• pnpm test (50 files, 460 tests)
• npm pack --dry-run --json

Notes

Live provider checks such as pnpm test:e2e, pnpm test:golden, and atlas-vision doctor remain maintainer-local because they require valid provider credentials and intentionally consume API spend.

atlas-vision-mcp v1.0.7

1.0.7 - Pi clipboard temp path fix

Fixed

• Auto-intercept now safely allows Pi CLI clipboard temp images such as pi-clipboard-*.png under the current OS temp directory.
• Fixes Image path is outside allowed directories when %TEMP% / %TMP% has been moved, for example to D:\Temp.
• Atlas-created attached-image and clipboard temp files are allowed per planned vision call without globally widening ATLAS_ALLOWED_DIRS; arbitrary temp image paths remain blocked.

Validation

• pnpm build
• pnpm lint
• pnpm typecheck
• pnpm vitest run tests/harness/intercept-images.test.ts
• pnpm test - 49 files, 444 tests
• npm pack --dry-run --json
• GitHub CI on main: green

atlas-vision-mcp v1.0.6

1.0.6 - Pi package gallery polish

Added

• Added a clean pi.image preview asset for https://pi.dev/packages/atlas-vision-mcp so the package renders with a clear, non-overflowing UI card in the Pi gallery.
• Expanded npm keywords with pi-extension, ocr, and screenshot for better Pi/npm discoverability.

Documentation

• Added a Pi-specific security note explaining local extension permissions, image and clipboard reads, provider upload behavior, and config review points.
• Clarified that pi install npm:atlas-vision-mcp is the supported distribution path; git install is not currently supported because the extension imports built files from the npm tarball.
• Converted README docs/examples links to GitHub absolute links so they work correctly from npm and Pi package pages.

Validation

• pnpm build
• pnpm lint
• pnpm typecheck
• pnpm test - 49 files, 442 tests
• npm pack --dry-run --json
• GitHub CI on main: green

atlas-vision-mcp v1.0.5

What's Changed

Fixed

• Added atlas-vision-mcp bin alias alongside atlas-vision, so npx atlas-vision-mcp@1.0.5 works for MCP client configs on Windows and other platforms.
• Clipboard image extraction is now cross-platform best-effort:
  • Windows: PowerShell Desktop Get-Clipboard -Format Image
  • macOS: pngpaste when installed, AppleScript fallback otherwise
  • Linux: wl-paste on Wayland or xclip on X11
• README/product docs now document clipboard platform support and backend requirements.

Validation

• pnpm lint
• pnpm typecheck
• pnpm test — 49 files / 442 tests
• pnpm build
• npm publish --dry-run
• Main branch CI green before tagging.

atlas-vision-mcp v1.0.4

What's Changed

Added

• Clipboard-first MCP tools: analyze_clipboard, ocr_clipboard, diagnose_clipboard, and analyze_ui_clipboard.
• OpenCode/Droid guidance for text-only models: copy an image to the OS clipboard, then ask Atlas to analyze the clipboard instead of using native image attachment.

Fixed

• Windows clipboard reads now run PowerShell in STA mode so Get-Clipboard -Format Image works reliably from MCP.
• serveStdio() now registers the same vision tools as the in-memory server, including the new clipboard tools.

Verified

• Local: pnpm lint, pnpm typecheck, pnpm test (49 files / 442 tests), pnpm build.
• E2E: OpenCode DeepSeek/GLM and Droid DeepSeek/GLM successfully called Atlas clipboard tools against a real Windows clipboard image.
• CI: main branch CI green before release.

atlas-vision-mcp v1.0.3

Fixes

• Gemini media_resolution enum format — mapDetailToMediaResolution() now uses correct proto enum names (MEDIA_RESOLUTION_LOW, MEDIA_RESOLUTION_ORIGINAL) instead of lowercase strings that were rejected by the API.
• Gemini model version gate — media_resolution is now only sent for Gemini 3+ models. Older models (gemini-2.x) skip the parameter entirely, avoiding Invalid value errors.

Added

• supportsMediaResolution(model) — exported helper to check Gemini model version compatibility.
• Gemini E2E tests — tests/e2e/gemini-e2e.test.ts (24 tests) covering all 7 tools against real Gemini models. Plus multi-model smoke suite for cross-version verification.

Validation

• pnpm typecheck ✅
• pnpm lint ✅
• pnpm test — 48 files, 434 tests ✅
• pnpm build ✅
• E2E verified on gemini-3.5-flash (17/24 pass, 7 quota-skipped)
• E2E verified on gemini-3.1-flash-lite (4/4 core tools pass)

atlas-vision-mcp v1.0.2

Fixed

• CLI eval now loads local .env values before resolving provider config, so pnpm test:golden works directly in local release checks.
• Synced CLI/package version metadata to 1.0.2.
• Refreshed release-readiness Harness evidence and golden snapshot baselines.

Validation

• pnpm typecheck
• pnpm lint
• pnpm test (48 files, 434 tests)
• pnpm build
• pnpm test:e2e (2 files, 32 tests)
• pnpm smoke:agents
• pnpm test:golden (16/16 snapshots, 0 gate failures)