mcp server
scopebond
Free, open-source guardrails and verifiable evidence for AI coding agents: Claude Code, Cursor, Codex, MCP, and GitHub. Scopebond Cloud is a separate proprietary hosted workspace.
Set up
Runs on your computer. Pick your agent and operating system.
Before you start
- Node.js 20 or newer
{
"mcpServers": {
"avouro-com-scopebond": {
"command": "npx",
"args": [
"-y",
"@scopebond/mcp"
]
}
}
}same format Cursor, GitHub Copilot CLI, Junie, JetBrains AI Assistant, Augment (Auggie CLI), Factory Droid, Kiro, Devin CLI +12
Values in angle brackets are placeholders. Restart Claude Desktop after editing its config.Claude Desktop docs
README
synced Oct 17, 2026Scopebond
See, control and prove what AI coding agents do on your computers.
Scopebond checks each action a coding agent asks to take (a shell command, a file read or write, a push, an MCP call) against your rules before it runs, and signs a record of every decision that you can verify offline. It works with Claude Code, Cursor, Codex, GitHub Copilot, MCP tools, agent frameworks and GitHub pull-request checks.
Get started · Explore the sample workspace · Docs
Get started
Windows
Create a free workspace, press Connect a computer and download the signed installer. Or run this in PowerShell:
irm https://cloud.scopebond.com/install.ps1 | iex
The installer is signed by Avouro LLC, needs no Node.js and keeps itself updated. It installs the Scopebond Agent (a companion program, not the AI coding agent doing the work), the hook, and the Scopebond tray. On first start the agent signs the computer in with a code you approve in the workspace and sets up the coding agents it finds. See Windows: the signed installer and the tray.
macOS and Linux
npx -y @scopebond/agent@latest setup https://cloud.scopebond.com
setup needs Node.js 22.13 or later. It signs the computer in with a code you approve
in the workspace, adds the hook to Claude Code (add --cursor or --codex for those
tools), and starts the Scopebond Agent with your sign-in. On a Windows computer that
uses npm, type npx.cmd; no execution-policy change is needed.
GitHub Copilot, or the hook alone
Run this in your own terminal, outside the coding agent:
npx -y @scopebond/hook@latest login https://cloud.scopebond.com --copilot
It prints a short code and an approval link for the workspace owner. Leave out
--copilot to connect Claude Code, or use --cursor or --codex. For Codex, review
Scopebond in /hooks, choose Trust, and start a new task. See the
hook README for each tool's details.
Local only, no account
npm install -g @scopebond/hook@latest
scopebond install
scopebond status
install configures Claude Code and detects Cursor, Codex and the GitHub Copilot CLI;
choose one with --cursor, --codex or --copilot. Records stay on this computer.
For one project's setup:
npx -y @scopebond/hook@latest init # Claude Code; add --cursor, --codex or --copilot
npx -y @scopebond/hook@latest log # inspect recent decisions
npx -y @scopebond/hook@latest verify # check stored records offline
In unattended setup, pass --yes to init after reviewing the configuration you are
authorizing.
Open source and hosted. Everything in this repository is free, Apache-2.0 open-source software, and it works without an account: the engine that checks and records every action, the starter protections, your own rules, and offline verification. Scopebond Cloud is a separate proprietary service with Free and paid plans. It adds curated rule packs that are kept up to date, Safe Install checks on the packages and downloads your agents install, one set of rules for every computer, approvals, inventory, and history for audits. Connecting one to the other changes the licence of neither.
How it works
- Every action is checked before it runs. The hook on the computer checks each action the coding agent routes through its tools against your rules, and decides there, without waiting for the workspace.
- Each rule is Watch, Needs approval or Block. In a workspace a rule is set for everyone, for one agent, or for a group of agents. Watch records the action, Needs approval stops it until it is unblocked, and Block stops it. Changing Scopebond's own settings is always blocked, so a coding agent cannot switch Scopebond off.
- Without an account, each rule records (the default) or blocks.
scopebond ruleslists them in plain words,scopebond rules enforce <rule>makes one block, andscopebond rules monitor <rule>sets it back to recording. The starter rules cover release branches, destructive programs, and secret, build and guardrail settings; network and MCP calls are recorded so you can decide which limits to apply. - Every decision is a signed record. Read them with
scopebond logand check them offline withscopebond verify.
When something is blocked
- On Windows, the tray shows it under Blocked · Allowed · Requests, in plain words, with the rule that stopped it. On macOS and Linux, and on Windows npm installs, the Scopebond Agent asks in its own window.
- What happens next follows the workspace's Who approves unblocking setting: An admin (the person asks to unblock it, with a reason, and an admin decides); Automatic, reviewed later (it is unblocked at once, for a length the workspace sets, and an admin reviews it later); or People unblock their own (the person unblocks it with a reason). A rule set to Block, and Scopebond's own protection, cannot be unblocked.
- Every unblock is recorded and shown in the workspace's Activity. Scopebond never runs the action itself; the person or the coding agent runs it again.
- Without an account, the coding agent is told which rule stopped it, and
scopebond ruleschanges what blocks.
The hosted workspace
Scopebond Cloud is a separate, proprietary service that brings every connected computer into one place:
- Overview: coverage (which computers are protected), what Needs you (requests to unblock, findings, rule reviews), and what was Stopped.
- Activity: every action and decision from every connected computer, unblocks included.
- Protection: the rules, for everyone, one agent or a group, and who approves unblocking.
- Share report: a summary for a period, as a share link or a PDF.
| The question | The open-source software | What the hosted workspace adds |
|---|---|---|
| What did an agent do? | Signed local records of supported actions and decisions. | Activity across every computer, and coverage on the Overview. |
| How are harmful actions stopped? | Rules that record or block supported actions before they run, or before merge. | Protection: Watch · Needs approval · Block for everyone, one agent or a group, who approves unblocking, and confirmation of which computers loaded each rule. |
| Can we prove it? | Signed records you can verify offline, without an account. | Retained records, Needs you, and a Share report. |
See workspace features and hosted plans.
The hook decides without waiting for the workspace to receive records. An action that needs approval still needs it: an unreachable workspace never grants permission. Without a workspace connection, records stay on the computer. Connected components send their records and health signals to the workspace.
Packages
| Package | Purpose |
|---|---|
@scopebond/hook | Checks for Claude Code, Cursor, Codex and GitHub Copilot, signed local records, and an optional workspace connection. |
@scopebond/agent | The Scopebond Agent: delivers records, keeps the connection and the hook working, and asks a person about a blocked action. Distinct from the AI coding agent doing the work. |
@scopebond/github-action | Required agent pull-request checks in your own runner. |
@scopebond/mcp | A policy-checking proxy for MCP tool calls. |
@scopebond/framework | Policy guards for tools inside agent applications. |
@scopebond/gateway | HTTP/MCP policy evaluation, constrained execution, signed evidence, and a kill switch. Default execution is a simulation. |
@scopebond/verify | Deterministic policy evaluation, signature verification, and conformance vectors. |
@scopebond/policy-schema | Policy, action, receipt, and observation schemas and shared vectors. |
@scopebond/sdk | Signing action intents and approvals and submitting them to a gateway. |
native | The signed single executable for Windows: the Scopebond Agent and the hook in one program, and its installer. Windows program, not on npm. |
tray | The Scopebond tray for Windows, a Tauri program. Windows program, not on npm. |
All of them are free to use and self-host under Apache-2.0. Not included, because they are part of Scopebond Cloud: the curated rule packs, Safe Install checks, the hosted portal, rules managed across many computers, approvals, inventory, shared retention and evidence, administration, the API and billing.
Coverage and limits
Supported tools
| Connection | What it checks | Important limit |
|---|---|---|
| Claude Code, Cursor, and Codex hook | Supported tool actions delivered through the configured hook. | Cursor file edits are recorded after writing, not prevented. Codex file reads are checked when derived from supported shell commands; reads outside that path are not seen. |
| GitHub Copilot hook | Shell commands, file reads and edits, and git pushes, checked before they run. The Copilot CLI is covered; VS Code agent mode and the Copilot cloud agent are handled, with verification on live sessions in progress. | Cooperative: prevents if honored, but a hook timeout fails open in Copilot, so the hook denies on its own shorter deadline. MCP tool calls are recorded, not evaluated. |
| GitHub Action | Policy checks on supported agent pull requests in your own runner. | Make the check required to gate merge. It does not monitor the agent's cloud sandbox; bypass rights remain outside the gate. Signing needs your configured signing key. |
| MCP proxy | Each tools/call routed through the proxy before forwarding. | Resource reads and prompt fetches are outside the tool-call policy. |
| Framework integrations | Guarded tools in Vercel AI SDK, LangGraph/LangChain, or your own tool loop. | The application must honor the guard; calls outside it are not covered. |
What to keep in mind
- Coverage depends on the tool and the connection. Only actions that reach a configured connection are checked; a process started outside the coding agent's tools is not covered. Some actions can be prevented; others are recorded after they happen. The hook capability manifest separates unsupported, inactive, configured, degraded, and verified-reporting states: a configured hook or a received signal alone is not verified coverage.
- A hook that cannot start fails open. Coding agents treat a hook that cannot
start as a non-blocking error and run the action unchecked.
scopebond statusandscopebond doctorreport a hook entry that cannot start. - The packages are experimental. Start with controlled test systems and review each package's coverage and operational limits before important work. Source availability is not a certification of production protection.
- A signature proves less than it may seem. It supports the integrity and provenance of what the signer asserted. By itself it does not prove an external effect, complete coverage, or regulatory compliance. See SPEC.md.
- Future connections are not product capabilities. Additional business-platform connections are ideas, not current features; see the product roadmap.
Verification and developer examples
SPEC.md describes evidence, action types, verification, and what a signature does and does not prove. Records omit file contents, prompts, and secrets; read package-specific data and redaction details before connecting.
Initialize a self-hosted gateway:
npx -y @scopebond/gateway@latest init
import { createGateway, StaticPrincipalKeyRegistry } from "@scopebond/gateway";
const keys = new StaticPrincipalKeyRegistry(principalKeyRecords);
const { app, handleAction } = createGateway({ policy, authentication: { keys } });
See the gateway README for simulation versus execution,
authentication, signing keys, and verification. Runnable examples live in
examples/; they do not require the proprietary hosted workspace.
Existing policy and verification conformance vectors are available here.
Releases
Current release set: @scopebond/policy-schema@0.11.0, @scopebond/verify@0.6.10,
@scopebond/gateway@0.20.4, @scopebond/sdk@0.1.11, @scopebond/hook@0.30.0,
@scopebond/agent@0.11.3, @scopebond/github-action@0.5.17, @scopebond/mcp@2.0.18,
and @scopebond/framework@0.3.18. Release tooling keeps these references in step
with package manifests. Public setup selects npm's latest published release;
generated hook configuration and managed workspace updates can pin versions
separately. The signed Windows programs are published on the
releases page.
Contributing and support
See CONTRIBUTING.md, SCOPE.md, and CODE_OF_CONDUCT.md; contributions are under a CLA. Report vulnerabilities privately as described in SECURITY.md. For hosted account or workspace help, use Scopebond support.
License
The software here is Apache-2.0 — see LICENSE and NOTICE. This license does not cover the proprietary Scopebond Cloud service. © 2026 Avouro LLC. Scopebond is a trademark of Avouro LLC.
Nothing in this repository is an offer of insurance, securities, or financial services, or legal or financial advice.