Skip to content

Releases: scy7796/visionsieve-mcp

Release list

VisionSieve MCP 1.0.0

VisionSieve MCP 1.0.0 Latest
Latest

Choose a tag to compare

View all tags
@github-actions github-actions released this 06 Aug 08:29
v1.0.0
7167348

VisionSieve MCP v1.0.0

VisionSieve gives text-first agents a narrow visual handoff: one focused image question goes to Luna, and compact checked evidence comes back to the host.

The default route uses the local Codex CLI, a ChatGPT login, and explicit gpt-5.6-luna. Verifier and fallback routes remain off unless the operator configures them. If the expected model or login route is unavailable, the request fails instead of switching to another model or an API-billed path.

New public identity

Version 1.0 introduces the VisionSieve name across the public interface:

  • Python distribution: visionsieve-mcp
  • CLI: visionsieve
  • MCP registration: visionsieve
  • configuration variable: VISIONSIEVE_CONFIG
  • standalone binaries and checksum manifest: visionsieve-*

The MCP tool remains vision.inspect.

The hardened runtime core stays compatible with 0.5 installations. Old visual-evidence-gateway console commands and the visual_evidence_gateway Python package remain available in 1.0 as migration aliases. Running visionsieve setup removes the old MCP registration before adding the new one.

Same-machine comparison

Six synthetic image tasks were run through native Codex image attachment and VisionSieve with the same questions.

Metric Native Codex attachment VisionSieve
Tasks completed 6/6 6/6
Expected fields fully present 4/6 6/6
Median visual text returned to the host 602 characters 62 characters
Median end-to-end time 16.6 s 20.2 s

On this fixture set, VisionSieve returned about one tenth as much visual text and kept every expected field. It was 3.6 seconds slower at the median. These six synthetic cases do not establish general accuracy, context-token, or speed results.

Installation

Prerequisite: Codex CLI signed in with ChatGPT.

Windows PowerShell:

irm https://raw.githubusercontent.com/scy7796/visual-evidence-gateway/main/install.ps1 | iex

macOS or Linux:

curl -fsSL https://raw.githubusercontent.com/scy7796/visual-evidence-gateway/main/install.sh | sh

The installer downloads a platform binary, requires a matching entry in visionsieve-SHA256SUMS.txt, registers the MCP server with an absolute command path, checks the ChatGPT login, and runs a small image probe. Setup failure rolls the new binary back.

Release assets

The release workflow builds:

  • Windows x86_64
  • Linux x86_64 and ARM64
  • macOS x86_64 and ARM64
  • wheel and sdist
  • source ZIP
  • SHA-256 manifest

Windows ARM64 has no prebuilt binary. Release binaries are not code-signed.

Boundaries

MCP protocol interoperability was validated over stdio with the official MCP SDK client in the 0.5 release process. Codex Desktop host discovery can vary by client version and may require a complete restart. Luna availability and latency depend on the account, region, workspace, and service conditions.

VisionSieve MCP is a community project, not an official OpenAI product.

Assets 11
Loading

v0.5.0

v0.5.0

Choose a tag to compare

View all tags
@github-actions github-actions released this 06 Aug 05:57
v0.5.0
2fa9a64

Release notes — Visual Evidence Gateway v0.5.0

What this is

A local, read-only MCP server that lets text-first agents (DeepSeek, OpenCode, Pi, …) call one tool, vision.inspect, for controlled visual evidence. The default backend is the local Codex CLI with explicit gpt-5.6-luna and forced ChatGPT subscription authentication; no API key, no silent fallback, no verifier by default.

Changes since vision-bridge-mcp v0.4.1

  • Renamed to Visual Evidence Gateway (visual-evidence-gateway distribution, visual_evidence_gateway package, visual-evidence-gateway CLI/MCP name, VISUAL_EVIDENCE_GATEWAY_* env vars). The old vision-bridge-mcp name collided with an existing same-purpose project on GitHub/npm; migration notes are in the README.
  • Transport-level model identity verification for the Codex CLI backend.
  • Hardened ChatGPT-mode environment stripping (OPENAI_BASE_URL, ANTHROPIC_API_KEY, GEMINI_API_KEY now also blocked).
  • Local-path masking in error/refusal output.
  • Windows junction/reparse paths are rejected cleanly (crash fixed).
  • Installers: mandatory SHA-256 verification, rollback on setup failure, explicit Windows ARM64 message.
  • Pre-release validation kit: strict JSON-Schema gate, cache-hit no-backend-call gate, junction negative test, fresh-cache runs, official-MCP-client host gate.
  • CI fixes: Windows wheel reinstall path, typer for the PyInstaller build.

Verified platforms

  • Windows 11 x64: full validation PASS (real Luna probes, six fixture types, security negatives, cache, schema, host MCP, installer simulation, PyInstaller single-file binary with real probe).
  • macOS/Linux: installer integration test + CI matrix (python 3.10–3.13); binary builds run in GitHub Actions (linux x86_64/arm64, macOS x86_64/arm64, Windows x86_64).

Not yet verified / boundaries

  • Codex Desktop app host call through MCP (manual restart + call needed; npm CLI 0.146.1 on Windows cannot complete stdio MCP tool calls — see FINAL_RELEASE_DECISION.md).
  • Windows ARM64 prebuilt binary (no GitHub-hosted runner).
  • Binary signing (none; transport + SHA-256 only).
  • Real-world latency on other accounts/regions (this machine: median ~21–25 s per visual call at validation time).
  • ModLens comparison run (requires external Antigravity CLI login).

Install

irm https://raw.githubusercontent.com/scy7796/visual-evidence-gateway/main/install.ps1 | iex
curl -fsSL https://raw.githubusercontent.com/scy7796/visual-evidence-gateway/main/install.sh | sh

Prerequisite: official Codex CLI (npm install -g @openai/codex) signed in with ChatGPT.

Loading