Keep AI approvals current

when anAI agentgains access.

Companies approve AI assistants for specific jobs. Later, someone quietly gives an assistant more access, to payments, code or customer data, and the approval no longer matches what it can do. Scopewake spots that change and tells the people who signed off.

Works with AI coding tools such as

  • Claude Code
  • Cursor
  • Windsurf
  • VS Code MCP
  • Claude Desktop
  • GitHub Actions
  • Gemini CLI
  • Codex CLI

In plain English

An employee is approved to read invoices. Months later, someone hands them the company card. Nobody updates the approval.

The same thing happens to AI agents. They are approved for one job, then quietly given more access as teams connect them to new tools. Scopewake is the record that notices.

  1. Approve

    A manager approves an AI assistant for one job, such as reading invoices. Scopewake records exactly what the assistant can reach on that day.

  2. Watch

    Engineers keep connecting assistants to new tools. Scopewake checks the setup files and notices when an assistant can reach something new.

  3. Re-check

    The approval is marked “needs review”, with the reason in plain words. The person who approved it decides what happens next.

Who it is for

Risk and compliance teams
Show that every AI approval still describes what the assistant can actually do.
Security engineers
See which AI assistants can reach payments, code, cloud accounts or secrets, and catch new access early.
Auditors and regulators
Get a dated record of each approval and every change since, in files they can test.

Words used on this page

AI agent
An AI assistant that can take actions, not just answer: run code, read files, call other systems.
Connector (MCP)
A plug that links an AI agent to another system, such as Stripe, GitHub or a database. MCP, the Model Context Protocol, is the common standard for these plugs.
Reach
Everything an agent could touch: its connectors, the passwords and keys it can read, and its cloud permissions.
CLI and CI
A CLI is a tool you run by typing a command. CI is the automated system that checks code before it ships; Scopewake can run there on every change.

Try it: change an agent’s access and watch the approval →

Change an agent’s access. Watch the approval.

Finance approved a Cursor agent on 1 October to read invoices. Its reach was frozen into the approval. Add or take away access and see what Scopewake reports.

Changes since approval

Rule: any addition to an approved agent’s reach needs review, whatever its severity. A removal alone does not.

Invoice reader

Approved by Head of Technology Risk, 2026-10-01 · owner Finance Operations

Re-review required
Scope
Read invoices, Singapore entity only
Agent
Cursor (project)
Valid until
2027-10-01
  • credential added generic-secret at .cursor/mcp.json:stripe (STRIPE_SECRET_KEY)
  • capability added payments via mcp:stripe
  • MCP server added stripe (local, npx)
$ scopewake review --ledger approvals.json

scopewake review (2026-10-24)

1 re-review required, 0 evidence expired, 0 to verify, 0 expiring soon, 0 current; 0 unapproved agent(s)

[re-review required] Invoice reader
  approved by Head of Technology Risk on 2026-10-01, valid until 2027-10-01; owner Finance Operations
  ! credential added: cursor/Cursor (project) (.cursor/mcp.json): generic-secret at .cursor/mcp.json:stripe (STRIPE_SECRET_KEY)  [MAS AIRM para 5.25(a), 5.24]
  ! capability added: cursor/Cursor (project) (.cursor/mcp.json): payments via mcp:stripe  [MAS AIRM para 5.25(a), 5.24]
  ! MCP server added: cursor/Cursor (project) (.cursor/mcp.json): stripe (local, npx)  [MAS AIRM para 5.25(a), 5.24]

How it works

Four moments in the life of one approval. Scroll and they stack, like the file they become.

An owner approves the agent.

scopewake approve records who approved what, for which agents, and freezes the reach they have today: servers, credentials, capabilities, cloud permissions.

Invoice reader

Approved 2026-10-01 · Finance Operations

  • invoices server
  • INVOICE_API_KEY

Reach frozen: 1 server, 1 credential

Its access grows without notice.

A developer connects Stripe to the same agent. Nothing in the approval record changes, so the approval now describes an agent that no longer exists.

Invoice reader

Approved 2026-10-01 · Finance Operations

  • invoices server
  • INVOICE_API_KEY
  • stripe server
  • payments capability

Change found: 1 server added

Scopewake flags the approval.

scopewake review compares today’s reach with the frozen snapshot. Any addition marks the approval re-review required, with each reason and the MAS paragraph it maps to.

Invoice readerRe-review required

Approved 2026-10-01 · Finance Operations

  • invoices server
  • INVOICE_API_KEY
  • stripe server
  • payments capability

3 reasons · MAS para 5.25(a)

The owner re-approves, on record.

The owner decides whether the change is significant and re-approves. The earlier approval stays in the history, so an examiner can see the whole trail.

Invoice readerCurrent

Re-approved 2026-10-24 · 1 earlier approval

  • invoices server
  • INVOICE_API_KEY
  • stripe server
  • payments capability

Reach frozen again, history kept

Sample evidence: review.md, review.csv, review.json

Open the folder. Click a paper to read it.

Evidence a reviewer can read

Every review names the approval, its owner and date, each change found, and the paragraph it maps to. One command, three formats.

$ scopewake review --ledger approvals.json --format md > review.md scopewake review --ledger approvals.json --format csv > review.csv scopewake review --ledger approvals.json --fail-on-review
Download the sample packRead the docs

Built for a careful yes

Start with local evidence.

It reads agent configuration files and policy exports on your own machines. Nothing is sent anywhere, so the evidence boundary is clear from the first day.

Show what is known, and what is not.

A configuration is evidence, not a guarantee of what an agent does at runtime. Gaps, unsupported sources and assumptions sit next to every finding.

Keep people in control.

Scopewake flags stale approvals and supplies the context. The accountable owner decides what is significant and what happens next.

Scopewake shows which approvals a change has touched, before anyone asks.

MAS Guidelines on AI Risk Management, para 5.25(a), issued 7 October 2026: significant changes should trigger appropriate reviews and re-approval processes before implementation. Scopewake is not endorsed by MAS; its mapping is indicative, not legal advice.

Scope

Agents it finds
  • Claude Codesettings, MCP
  • Cursor, WindsurfMCP config
  • VS Code MCP, Claude DesktopMCP config
  • Gemini CLI, Codex CLIdetected
  • GitHub Actionsworkflows that run agents
agents: 3 found
  cursor/Cursor (project)        score 64
  claude-code/Claude Code        score 41
  github-actions/review.yml      score 22
Credentials it lists
  • MCP server env and argsnames only
  • .env files in the projectnames only
  • Cloud and tool files in homeAWS, GCP, Azure, kube, SSH
  • Workflow secretssecrets.* references
credential: generic-secret
  at .cursor/mcp.json:stripe
  name STRIPE_SECRET_KEY
  value never read into output
Capabilities it infers
  • Shell and filesystem
  • Browser control
  • GitHub, Slack, email
  • Databases, cloud, paymentsfrom MCP server names
capability: payments via mcp:stripe
capability: github via mcp:github
capability: shell-execution via cursor
Cloud permissions
  • AWSfrom an IAM export you supply
  • Role chainsup to three AssumeRole hops
  • GCP and Azurenot analysed yet
$ scopewake aws iam.json
  ci-deployer      privilege escalation
  dev-agent        data read, secrets read
What it does not do
  • Block or enforceit reports only
  • Call cloud APIsnever
  • Verify keys are liveno
  • Certify complianceno; owners decide
Indicative mapping, not legal advice.
The owner decides whether a change
is significant.

Start with one assessment

The scanner is open source and free to run on your machines and in CI. You pay when you want it set up, kept current and turned into evidence.

S$one-off
  • Your agents’ reach scanned on your own machines
  • Each AI approval checked against today’s reach
  • A findings report with the evidence files attached
  • Scope and acceptance criteria agreed in writing first

Then an annual evidence programme, from S$15,000.

Questions buyers ask

Scopewake is built to open no network connection and write no files. It reports the names, kinds and locations of secrets, never their values. Tests and a source-policy check enforce this design; they are not a proof, so read the code and run it offline first.

Start with one agent

Define a small pilot your team can evaluate in a week. Choose who leads it and where the first evidence comes from, and take the brief to your next meeting.

Scopewake pilot brief

Team: Technology risk
Agent workflow: (name one agent workflow)
First evidence source: MCP connector configuration

Plan
1. Record the approved baseline for this one agent with `scopewake approve`.
2. Name the accountable reviewer and the approval owner.
3. Introduce one known access change in a test configuration.
4. Run `scopewake review` and export the result (Markdown, CSV or JSON).
5. Decide together whether the evidence supports a useful review.

Out of scope: runtime enforcement, certification, legal opinion.
Reference: MAS Guidelines on AI Risk Management, para 5.25(a) (indicative).

Created in your browser. Nothing was submitted.

Created in your browser. Nothing is submitted.

Ask about an assessment

This opens a ready-written email in your own email app. Nothing is sent from this page.

Prefer to read first? Download the sample evidence or build a pilot brief.