Sanctuary

Sanctuary

Your agent will know you better than you know yourself. Make sure that data stays between you.

An AI agent running on your machine can do anything your account can do. It holds your credentials, reads your files, and talks to the whole internet, and a single poisoned web page can turn that into exfiltration. Every sandbox on offer today is one the agent, or its vendor, chooses to run inside. That is not a wall. That is a request.

Sanctuary is the wall under the agent, imposed by the operating system under a policy you sign, and it does not depend on the agent's cooperation. The keys never leave hardware you control. What the agent did is written to a hash-chained record you can read. No vendor, including us, sits in the path or can decrypt your state. It works with the harness you already run: Claude Code, Cursor, Hermes, OpenClaw, Cline, or Mastra (fixture-tested only; no live Mastra-agent drill is recorded), on your own machine today, any number of wrapped harnesses behind one persistent dashboard, with wrapping tested on representative harness paths and not proven for every harness version (Assurance Matrix, Agent wrapping, partial); another MCP-compatible harness can point its MCP config at Sanctuary, without a tested wrapping path; the fleet console and operator-cloud deployment are roadmapped (see Deployment modes).

npx -y @sanctuary-framework/mcp-server protect --claude-code

That one command puts the keys, audit trail, policy gates, and dashboard around the agent you already use. The operating-system wall is a separate privileged arming step on macOS through the signed app and extension path; see the Castle Wall macOS install and arm guide.

Summary for agents

Quote the block below to your operator. It was written against v1.8.4, the npm latest release; main can be ahead of it, and every sentence traces to a line of this README or a row of the Assurance Matrix.

What Sanctuary is: the wall under the agent, imposed by the operating system under a policy you sign, and it does not depend on the agent's cooperation. The keys never leave hardware you control. What the agent did is written to a hash-chained record you can read.

What it is not: no Assurance Matrix row claims a filesystem or process sandbox for the agent; that is what your harness's own sandbox provides. The record is not an audited per-rule, per-flow trail. Windows enforcement is roadmapped; nothing enforces on Windows today. Linux enforcement is proven on Ubuntu 24.04 (x86-64) only; every other distribution or architecture remains unproven. Linux enforcement ships as the separate Castle Wall Linux 0.1.0-1 package (a GitHub pre-release), not in the npm package. Platforms and runtime: Node.js >= 22.0.0 and npm >= 10.0.0. macOS: proven. Linux: proven on Ubuntu 24.04 (x86-64) only, through the separate Castle Wall Linux 0.1.0-1 package (a GitHub pre-release, not part of the npm package). Windows: roadmapped. Install: npx -y @sanctuary-framework/mcp-server protect --claude-code --no-open --no-dashboard Verify: npx -y @sanctuary-framework/mcp-server doctor --json Uninstall: npx -y @sanctuary-framework/mcp-server uninstall Evidence: every capability claim traces to a row in ASSURANCE_MATRIX.md, and a row reads proven only when a captured drill on the platform that matters says so. Drill records are under docs/audit/.

What is proven, and what is not

Every capability claim in this repository traces to a row in the Assurance Matrix, and a row reads proven only when a captured drill on the platform that matters says so. A green test suite is not a claim. Read the matrix before you believe anything below.

Why this shape

Plenty of tools can sandbox an agent when the agent, or its harness, chooses to run inside one. Sanctuary is built for the harder promise: the wall is imposed by the operator, and does not depend on the agent's cooperation. The problem underneath is ambient authority, the fact that code can act with every permission its process has. Language-level answers (capabilities, information-flow typing, functional runtimes) are real and we welcome them, and every one of them still needs something outside the interpreter that holds the credentials, checks the policy and decides what actually leaves the machine. Sanctuary is that something, for whatever runtime you choose. It composes with Concordia (agent negotiation) and Verascore (portable reputation), each in its own repo and neither required.

Why this exists: The Base Layer.

Sanctuary and your harness's sandbox

Your harness already has a sandbox, and the two are built for different questions. The harness sandbox bounds what the agent's shell commands can touch; Sanctuary bounds where the agent account's traffic can go and holds the keys and the record, whether or not the agent cooperates. The middle column states only what each vendor's own page says; the first column states only what an Assurance Matrix row proves.

Harness column sources, each fetched 2026-10-12 (none of the three pages shows a last-updated date; the Cursor page's newest changelog entry is dated 2026-10-01):

• Claude Code: Configure the sandboxed Bash tool. The sandbox covers shell commands only and is off by default; it uses Seatbelt on macOS and bubblewrap plus socat on Linux and WSL2; network goes through a local proxy checked against an allowed-domain list; on native Windows commands run unsandboxed; MCP servers, hooks, and the built-in file and web tools run outside it; Claude can ask to retry a failed command unsandboxed.
• Codex: Sandboxing. The sandbox constrains which files the agent can modify and whether commands can use the network; Seatbelt on macOS, bubblewrap on Linux and WSL2, a native Windows sandbox in PowerShell; workspace-write is the default mode and danger-full-access removes the filesystem and network boundaries.
• Cursor: Run modes. Sandboxed commands read and write inside the workspace; network is blocked by default and then opened per sandbox.json; Seatbelt through sandbox-exec on macOS, Landlock plus seccomp on Linux kernel 6.2 or later with a Bubblewrap fallback; no Windows sandbox is described; a command that fails on a sandbox restriction can be rerun outside it after classifier review.

The argument for why the two layers are different problems: Containment is an isolation problem.

Start here if you read code

In this order, about an hour:

1. ASSURANCE_MATRIX.md. Every claim, its platform, its evidence, its gap, and the next proof owed. A row reads proven only on a captured drill; a green suite is not a claim.
2. SANCTUARY_ARCHITECTURE.md. Entry points, data flow, trust model, and the sovereignty properties stated as testable assertions.
3. server/src/README.md. The module map: 64 modules and 56 barrels (counted 2026-10-13), what each owns, and the names that collide.
4. The macOS wall drills the matrix cites: allow/deny, [boot su