Panel Review

ai.truverif由 truverif.ai 發布官方發布者v0.19.50更新於 Nov 10, 2026

Guardian agent for AI coding: four frontier models review risky diffs and commits before they ship.

已驗證Streamable HTTP可網頁執行開發者工具AI 與機器學習

概覽

AI 產生的概覽

讓 AI 程式助理把差異與提交交給多模型評審面板,並在本機執行提交前評審閘門。

可以做什麼
Panel Review 是遠端 MCP 伺服器,為程式代理提供多模型評審工具(例如 audit_coding),分析代理明確提交的差異或上下文。搭配的安裝程式會加入本機提交前與寫入前評審閘門,在 Claude Code、Codex CLI、Cursor、VS Code/Copilot、Gemini CLI 與 Antigravity 中攔截有風險的提交與編輯。閘門採失敗放行:服務無法連線時,提交或寫入照常進行並顯示提示。
適用情境
當你希望在程式變更發布前取得自動化的第二意見,且使用受支援的代理主機時,值得加入。適合想對提交與代理編輯設置評審閘門、而不只依賴人工評審的團隊。
執行需求
透過 streamable HTTP 連線遠端端點 Authorization 標頭攜帶 Bearer API 金鑰(來自提供者設定頁面)。安裝程式需要機器上有 Node 18+ 與 Python 3,並需要網路存取提供者 API 以完成登入與閘門檢查。
安裝前請注意
Authorization 標頭攜帶 Bearer API 金鑰;安裝程式會把 tvai_ 金鑰寫入本機設定檔,並可能對代理與儲存庫設定寫入鉤子與 MCP 項目。閘門檢查會傳送雜湊後的儲存庫指紋、變更片段內容雜湊,以及本機分類器的類別標籤與分數給提供者,但不傳送原始碼與檔案路徑。評審工具會傳送代理傳入的差異或上下文。解除安裝會在伺服器端撤銷金鑰,若仍殘留含密鑰的內容則以非零狀態結束。

安裝

在 SourceWeft 中

  1. 開啟 儀表板中的 Panel Review,將其新增到工作區。
  2. 連接 Authorization token 憑證。它們會被私密儲存,絕不會在公開頁面上展示。
  3. 為需要使用其工具的對話啟用該服務。

Web executable,透過 Streamable HTTP。 遠端服務在工作區中設定後即可從網頁執行環境執行。

交給代理程式安裝

把這段提示詞貼上到 Claude Code、Codex、Cursor 等代理程式,或 SourceWeft 對話中。代理程式會讀取該伺服器的安裝指南,先向你展示發布者、端點和登入方式,經你同意後再把它新增到你的客戶端。金鑰不會寫進任何命令。

閱讀 https://sourceweft.com/mcp/ai.truverif%2Fpanel-review/install.md,並按其中說明把這個 MCP 伺服器接入我的客戶端。修改任何設定之前,先告訴我它的發布者、端點和登入方式,等我確認。不要把 API 金鑰寫進命令或我會提交的檔案;問我金鑰應從哪裡取得。

檢視代理程式所依據的安裝指南

其他 MCP 客戶端

把它新增到你客戶端的 mcpServers 設定中。

{
  "mcpServers": {
    "panel-review": {
      "type": "http",
      "url": "https://mcp.truverif.ai/mcp",
      "headers": {
        "Authorization": "Bearer <YOUR_TOKEN>"
      }
    }
  }
}

README

@truverifai/init

One command to connect your AI coding agents to TruVerifAI — multi-model review tools plus local pre-commit/pre-write review gates — on Claude Code, Codex CLI, Cursor (IDE + CLI), VS Code / GitHub Copilot, Gemini CLI, and Antigravity.

[The write gate blocking a risky auth change, live]

npx @truverifai/init          # plan -> one confirm -> browser login -> install gates + tools -> verifynpx @truverifai/init doctor   # re-verify anytime: gates armed, tools connected, key validnpx @truverifai/init logout   # remove the API key from every config this tool wrote

init shows a plan of everything it will write for your machine and asks once before touching anything. To scope or preview it:

npx @truverifai/init --dry-run              # print the plan, write nothingnpx @truverifai/init --only claude,codex    # install for these agents onlynpx @truverifai/init --skip gemini,hook,rules   # everything except thesenpx @truverifai/init --yes                  # non-interactive (CI): accept the plannpx @truverifai/init --version              # print the version, do nothing else

Agent names: claude, codex, copilot, vscode, cursor, gemini, antigravity; hook is the git pre-commit gate and rules the agent-rules blocks. Declining the prompt exits with nothing written. A flag init does not recognize also exits with nothing written (it never falls through to an install).

MIT-licensed. Zero runtime dependencies — this package is plain, unminified JavaScript and Python; npm pack @truverifai/init and read every line. There is no build step and no transitive supply chain.

See it in action

Requirements

  • Node 18+ — runs this installer, and every hook command it writes launches a gate via node <launcher> <host> <gate>.py. No node, no gates, on any host.
  • Python 3 (python3, or py on Windows) — the gate code itself.

npx @truverifai/init doctor checks both and names the missing one, rather than letting an absent runtime disable the gates silently.

Turning the gates off (and back on)

The review gates block risky commits and edits. One command controls all of them:

bash
npx @truverifai/init gates off       # stop every gate, everywherenpx @truverifai/init gates on        # turn them back onnpx @truverifai/init gates status    # what's on, and which setting decided it

(If you installed globally with npm i -g @truverifai/init, the same commands are available as the shorter tvai gates off. Plain npx does not leave a tvai binary on your PATH, so use the full form above unless you installed globally.)

This writes enable_gates to ~/.truverifai/config.json, which every delivery reads — the git pre-commit hook and the Codex, Cursor, Copilot, VS Code, Gemini and Antigravity hooks alike. The MCP review tools stay connected either way; only the automatic gating stops.

Two things worth knowing:

  • Claude Code no longer has its own toggle. enable_gates was removed from the plugin's settings panel (it could only ever reach Claude Code's own hooks — the X8 split), so this command is the single switch. One exception: a value stored in the panel before the removal may still be exported to Claude Code's hooks until you clear it; gates status calls that out and says how.
  • TVAI_ENABLE_GATES overrides everything. If you have it exported, gates on|off writes the file underneath it and has no effect until you unset it. gates status says so.

To remove the gates entirely rather than switch them off, see Removing it below.

What this installs, exactly

init runs as you, interactively: it prints the plan first, writes only after you confirm, and prints every file it touches. The complete list:

Its own home — ~/.truverifai/

  • config.json — your tvai_… API key (minted via browser device-flow login; this process never sees a password).
  • gates/current/ — the review-gate code (Python + a Node launcher), vendored from this package's vendor/ directory.
  • .nudge_state.json, gate_state/ — local rate-limit / state files.

Per detected agent (user-level; only for agents found on your machine)

  • Claude Code: the plugin's api_token option, and one permissions.allow entry (mcp__plugin_panel-review_truverifai) in ~/.claude/settings.json so Claude's auto-mode classifier permits the free gate-release calls. Additive merges only — the file is backed up first, never created, and never has anything removed.
  • Codex CLI: an MCP server block (between TRUVERIFAI markers) and [features] hooks = true in ~/.codex/config.toml; ~/.codex/hooks.json.
  • Cursor: ~/.cursor/hooks.json (gates) and ~/.cursor/mcp.json (tools).
  • Copilot CLI: ~/.copilot/mcp-config.json.
  • VS Code: your user mcp.json.
  • Gemini CLI: ~/.gemini/settings.json (tools entry).

Per repository (written in the repo you run init from, with a prompt)

  • .github/hooks/truverifai-gate.json + truverifai-vscode.json (Copilot CLI / VS Code gate hooks), .gemini/settings.json (Gemini hooks), .agents/hooks.json (Antigravity hooks).
  • Optional agent-rules blocks (marked TRUVERIFAI_RULES_START…END) in AGENTS.md / CLAUDE.md / GEMINI.md — you are asked first.
  • .git/hooks/pre-commit — the git gate, which catches commits made outside any agent (and is the one layer --no-verify makes an explicit, auditable act). Installed automatically when you run init inside a git repo. It is the only repo-scoped gate, so add it to other repos with cd <repo> && npx @truverifai/init hook. An existing pre-commit hook you wrote is never overwritten — init reports it and prints the line to add by hand.

logout strips the API key from every config listed above. Nothing is installed as a service, daemon, or startup item; the gates only run when your agent host invokes its hooks.

What leaves your machine

  • Login: a device-flow handshake with truverif.ai mints your API key in the browser.
  • Gate checks (on risky writes/commits): the gates POST to api.truverif.ai a hashed repo fingerprint (SHA-256 of your git remote URL or repo path — the raw path/URL never leaves the machine), content hashes of the changed hunks, and the local classifier's category labels and scores. Not your source code, and not your file paths (only coarse path-class tags like "test/docs").
  • Review tools (audit_coding etc.): send only what your agent explicitly passes when it invokes a review — that is the product: the diff/context you choose to submit is analyzed by multiple frontier models.
  • Update check: gate responses may include a "newer version available" note computed server-side against the public npm registry; the client sends only its own version string.
  • Diagnostics (TVAI_PAYLOAD_LOG) are local-only and opt-in — nothing is uploaded.

The gates fail open by design: if our server is unreachable or anything errors, your commit/write proceeds and a visible notice says the change was not gated. This tool never blocks your work on its own failure.

Verifying this package

The full source is public at github.com/TruVerifAI/init — every release is synced there. To verify a tarball by hand: npm pack @truverifai/init, extract, and diff against the repo — bin/tvai.js, lib/*.js, and vendor/gates/*.py are the entire runtime surface, dependency-free. The gate code (vendor/gates/) is also published at github.com/TruVerifAI/claude-plugins (plugins/panel-review/hooks/). Build provenance attestation (CI publish with cryptographic linkage to this repo) is the planned next step.

Removing it

bash
npx @truverifai/init uninstall

A complete removal, in one command:

  • Every hook config init wrote (Codex, Cursor, Copilot, VS Code, Gemini, Antigravity, and the git pre-commit hook), the vendored gate code under ~/.truverifai/, and the MCP server entries. Hook files you share with other tools are edited, not deleted — only our own entries are taken out, matched by marker.
  • Your API key, revoked server-side. Before deleting the local key file, uninstall calls the API to revoke the key itself, so any leaked or synced copy of it is dead. If the server is unreachable it says so and continues (the local removal never blocks on the network).
  • The plugin token in Claude's credential store — on a Mac this is the Keychain entry's panel-review value, edited in place (the Keychain item itself is Claude's and is left alone).
  • The .tvai-bak backups init made of shared config files — deleted only after the live file is verified clean; if a live file still contains our entries, its backup is kept and named as the restore source.

Uninstall is honest about failure: anything it could not remove is listed at the end under "needs your attention", and the command exits nonzero when secret-bearing residue remains (CI scripts that assumed exit 0 should check this). Silence means clean.

Uninstall also tries the host CLIs to remove the plugins init installed (claude plugin uninstall panel-review@truverifai, codex plugin remove panel-review@truverifai), best-effort: each attempt prints what happened, a failure never blocks the rest of the uninstall, and where no runnable CLI exists (the Claude desktop app) it prints the in-app step instead (+ button, then Plugins).

npx @truverifai/init logout is the lighter option: it clears the key (and our backups of files that held it) and the MCP entries but leaves the hooks installed. They then fail open for want of a key, and a later npx @truverifai/init login re-arms them.

Support

來源:README.md,提交 3192e15

工具

0
該伺服器需要 API 金鑰等憑證,因此 SourceWeft 無法在沒有憑證的情況下讀取其工具,也未收錄。

版本歷史

6
  1. v0.19.50最新Oct 13, 2026
  2. v0.19.49Sep 23, 2026
  3. v0.19.48Sep 23, 2026
  4. v0.19.47Sep 18, 2026
  5. v0.19.46Sep 17, 2026
  6. v0.19.45Sep 16, 2026

評分

已滿足適用檢查項的 85%

該評分衡量此收錄條目的完整程度,依據註冊中心條目、端點與公開來源計算得出。它不是對伺服器本身的評價,也不表示伺服器安全或目前可用:可連通性是帶日期的檢查結果,評分也不包含安裝量或使用者評價。

  • 近 30 天內可連通必需25 分
  • 已宣告認證方式必需10 分
  • 已列出工具15 分

    請讓伺服器在無憑證時回應 tools/list,或發布可在沙箱中無憑證啟動的軟體包。

  • 描述必需15 分
  • 圖示10 分
  • 文件15 分
  • 授權條款10 分