Fabric Verify

ai.augmentev由 augmentev.ai 發布v0.3.1更新於 Nov 3, 2026

Signed, publicly logged records of what your AI agent did, with proofs anyone can check offline.

發布方自己的服務:Fabric

已驗證Streamable HTTP可網頁執行開發者工具安全與監控

概覽

AI 產生的概覽

把 AI 代理的動作簽章記錄到防竄改公開日誌,並產生可離線驗證的證明。

可以做什麼
  • 記錄動作與結構化代理事件(工具呼叫、結果、核准、委派、政策決定、中止等)並寫入簽章公開日誌
  • 取得並儲存已記錄項目的公開證明,匯出事件包供離線驗證
  • 查看剩餘預付額度與本次工作階段的花費
  • 產生購買額度的信用卡結帳連結
適用情境
當代理會執行有後果的步驟——決策、付款、核准、部署、資料寫入——而你需要一份第三方或稽核人員不必信任你或廠商即可離線驗證的防竄改紀錄時適用。
執行需求
使用遠端端點,或在 Node.js 20 以上執行 npm 套件 @augmentev/fabric-verify-mcp。需要預付額度金鑰 FABRIC_CREDIT_KEY(fvk_...)、代理名稱 FABRIC_AGENT,可選 FABRIC_KID、FABRIC_HOME、FABRIC_MAX_SPEND_USD。遠端端點宣告使用 OAuth。需能連線至 Fabric Verify 服務。
安裝前請注意
每筆記錄從預付額度扣 0.005 美元,單一工作階段上限由 FABRIC_MAX_SPEND_USD 控制(預設 1);額度不可退款,餘額上限 1000 美元。請妥善保管 FABRIC_CREDIT_KEY。記錄永久保存並以指紋形式公開。代理與金鑰名稱由廠商保存,切勿在 FABRIC_AGENT 或 FABRIC_KID 填入個人資料。匯出的事件包會向收件者揭露其內容。請備份代理金鑰,遺失後須更換 FABRIC_KID。

安裝

在 SourceWeft 中

  1. 開啟 儀表板中的 Fabric Verify,將其新增到工作區。
  2. 連接 所需的 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.augmentev%2Ffabric-verify/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。 客戶端會在首次連接時完成 OAuth 登入。

{
  "mcpServers": {
    "fabric-verify": {
      "type": "http",
      "url": "https://mcp.augmentev.com/mcp"
    }
  }
}

README

Fabric Verify MCP server

Give your AI agent a tamper-evident record of what it did.

Each time the agent records an action, this server signs a short record on your machine, with the agent's own key, and adds it to the Fabric Verify public log. Later you get a proof that anyone can check offline, without trusting you or us.

A proof shows that this exact record existed, unchanged, by a certain time. It does not show that what the record describes is true.

Your text stays with you. Each record sends a salted fingerprint of it, plus the signature. Action details stay private. The public log holds only a salted fingerprint, the action kind and time, and coded identifiers that only you, or an auditor you give an audit pack to, can link to your agents. Records are permanent. AugmentEV keeps the agent name, key name and public key in an erasable store so they can go into your audit packs, so never put personal data in FABRIC_AGENT or FABRIC_KID.

Price

ToolCost
record_action, record_event$0.005 per record, from prepaid credit
get_proof, check_balance, buy_credit_link, export_eventsfree

Buy credit by card at https://augmentev.com/pricing#buy. Your credit key (fvk_…) is shown once after payment, so save it in a password manager.

Install

You need Node.js 20 or later. Add this to your MCP client's configuration:

json
{  "mcpServers": {    "fabric-verify": {      "command": "npx",      "args": ["-y", "@augmentev/fabric-verify-mcp"],      "env": {        "FABRIC_CREDIT_KEY": "fvk_...",        "FABRIC_AGENT": "my-agent"      }    }  }}

In Claude Code:

bash
claude mcp add fabric-verify --env FABRIC_CREDIT_KEY=fvk_... --env FABRIC_AGENT=my-agent -- npx -y @augmentev/fabric-verify-mcp

Settings

VariableDefaultMeaning
FABRIC_CREDIT_KEY(required)Your credit key. Keep it secret.
FABRIC_AGENTmcp-agentA stable name for this agent (printable ASCII, no spaces).
FABRIC_KID<agent>-key-1A name for the agent's signing key. Change it when you rotate the key.
FABRIC_HOME~/.fabric-verifyWhere the agent key and your records are kept (private to your user).
FABRIC_MAX_SPEND_USD1The most this server may spend in one session.

Tools

  • record_action (kind, summary, optional idempotency_key, optional approval_record): signs and logs one record, then returns its log index. The summary and its salt are saved under FABRIC_HOME/records/<api host>/<index>/, so you can later show what the record was about. If a call fails, call again with the same idempotency_key, kind and summary: the exact same signed record is resent and charged at most once.
  • record_event (event, body, optional action, run_id, step_id, parent_record, delegation_record, kind, idempotency_key): signs and logs one structured agent event: a tool call, its result, an approval, a delegation, a policy decision, a halt and so on (format augmentev/fabric-agent-event/v1). Same price as record_action. The event is the record's payload and stays on this machine. It returns record, which later events link to. For a tool call, pass action (tool, server, arguments), and pass the same action when recording its approval, so both name the same action. If your account's governance policy requires approval for tool calls, set body.approval_record and the approval is presented to the log, which uses it once. A tool result is signed as kind other, so that rule does not also gate results.
  • export_events (run_id or indices): bundles events recorded from this machine, with their proofs, into one file that anyone can check offline with augmentev-verify agent-events. Free. Fetch the proofs first. The bundle shows that its events were logged unchanged and that every link points to an earlier event of the right type. It does not show that the bundle holds every event, or that a described tool result is true. Events recorded on another machine (another agent's approval, say) are in that machine's bundle; merge records in log order to check them together.
  • get_proof (index, optional wait_seconds up to 45): fetches the proof and saves it next to the record. Proofs are usually ready a few minutes after recording.
  • check_balance: shows your remaining credit and this session's spend.
  • buy_credit_link (amount_usd: 25, 50, 100 or 500): returns a card checkout link for a person to open. This tool charges nothing.

Check a proof

Download the verifier and the published log keys from https://augmentev.com/verify, then run this in the record's folder:

bash
augmentev-verify log-proof proof.json --pins fabric-log-v1.json --audit agent-record.json

Pin the version

For production agents, pin an exact version, such as npx -y @augmentev/[email protected], and update deliberately. The package ships a lockfile (npm-shrinkwrap.json) so its dependencies are fixed too.

Keep your agent key

The agent key is FABRIC_HOME/agent-ed25519.pem. Back it up. Each key name (FABRIC_KID) belongs to one key for good. If you lose the key, set a new FABRIC_KID.

Support

[email protected]

Licence

Apache-2.0. "Fabric Verify" and "AugmentEV" are trademarks of AugmentEV Inc.; the licence grants no rights to them.

來源:README.md,提交 3e39ce1

工具

6

record_action

Record an agent action (costs $0.005)

Sign a record of an action this agent took and add it to the Fabric Verify public log. Spends $0.005 of prepaid credit. The summary stays on this machine. The public log permanently holds only a salted SHA-256 fingerprint of the summary, the action kind and time, and coded identifiers that only the user, or an auditor they give an audit pack to, can link to this agent. Returns the record's log index; fetch its proof later with get_proof. A proof shows that this exact record existed, unchanged, by a certain time. It does not show that what the record describes is true.

輸入參數kind必填summary必填approval_recordidempotency_key

record_event

Record an agent event (costs $0.005)

Sign a structured record of one governed step and add it to the Fabric Verify log: a tool call, its result, an approval, a delegation, a policy decision, a halt, and so on (format augmentev/fabric-agent-event/v1). Spends $0.005 of prepaid credit. The event stays on this machine with its salt. Action details stay private: the public log holds only a salted fingerprint of the payload, the action kind and time, and coded identifiers that only the user, or an auditor they give an audit pack to, can link to this agent. Returns `record`: later events link to it (a tool call's approval_record, a tool result's request_record, parent_record). For a tool call, give `body.tool` and `body.arguments` (hashed for you), or `action`; an approval of that call gives the same `action`, so both name the same action. When the account's policy requires approval for tool calls, set body.approval_record and it is presented to the log. Bundle a run with export_events for an offline check. The check shows each event was logged unchanged and every link points to an earlier event of the right type. It does not show that a described tool result is true.

輸入參數body必填kindevent必填actionrun_idstep_idparent_recordidempotency_keydelegation_record

export_events

Bundle agent events for an offline check

Gather agent events recorded from this machine, with their proofs and salts, into one file (augmentev/fabric-agent-events/v1) that anyone can check offline with `augmentev-verify agent-events`. Free. Fetch each event's proof with get_proof first. The file discloses the events' contents to whoever receives it. It shows the bundled events, not that the bundle holds every event: completeness comes from the account's governance export.

輸入參數run_idindices

get_proof

Get the proof for a record

Fetch the public proof for a recorded action and save it next to the record, with the agent record needed to check it. Free. Returns "pending" if the record is not in a public checkpoint yet (usually a few minutes). Anyone can check a saved proof offline with the verifier at https://augmentev.com/verify. A proof shows that this exact record existed, unchanged, by a certain time. It does not show that what the record describes is true.

輸入參數index必填wait_seconds

check_balance

Check Fabric Verify credit

Show the remaining prepaid credit, about how many records it covers, and this session's spend against its limit. Free.

無需輸入參數。

版本歷史

5
  1. v0.3.1最新Nov 1, 2026
  2. v0.3.0Nov 1, 2026
  3. v0.2.1Oct 23, 2026
  4. v0.2.0Oct 21, 2026
  5. v0.1.0Oct 21, 2026

評分

已滿足適用檢查項的 100%

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

其他 Fabric 整合