Skip to main content
Glama

CesSpace ARC — Secure Agent-to-Machine Control Plane

License Stage Security Policy Engineering Governance Node pnpm

CesSpace ARC (Agent Remote Control) is a secure, vendor-neutral control plane providing policy-enforced, audited access to development machines, virtual machines, repositories, terminals, and processes for authorized AI coding agents. Lead Developer & Architect: P Thakur


1. Mission & Vision

Modern AI coding agents (Claude Code, Antigravity, OpenAI Codex, OpenCode, DeepSeek agents) require rich interactions with host environments—reading files, inspecting diffs, running tests, and executing builds. However, giving autonomous agents unrestricted ambient access (such as open SSH keys, raw /bin/sh evaluation, or unfettered root access) creates severe security risks:

  • Accidental or malicious filesystem destruction.

  • Credential harvesting (~/.ssh, ~/.aws, .env).

  • Indirect prompt injection leading to remote code execution.

  • Silent repository or supply chain tampering.

CesSpace ARC mediates every single agent action through an authoritative policy engine, an OS-level filesystem jail, and an append-only audit trail.


Related MCP server: MCP SSH Orchestrator

2. High-Level Target Architecture

AI Client / Coding Agent (Claude Code / Antigravity / Codex / DeepSeek)
                          │
                          ▼
             MCP Interface (JSON-RPC 2.0)
                          │
     ═════════════════════╪═══════════════════════ Trust Boundary 1 (Ingress)
                          ▼
                 ARC Remote Gateway
                & Authentication Engine
                          │
     ═════════════════════╪═══════════════════════ Trust Boundary 2 (Policy)
                          ▼
                    Policy Engine
            (RC-01: Minimal Security Kernel;
             RC-04: Declarative Rules & Approvals;
             Precedence: DENY > REQUIRE APPROVAL > ALLOW)
                          │
                          ▼
                     Audit Layer
            (RC-01: In-Memory / Stream Sink;
             RC-06: Anchored Persistent Log)
                          │
     ═════════════════════╪═══════════════════════ Trust Boundary 3 (Host Execution)
                          ▼
                    ARC VM Agent
                          │
         ┌────────────────┼────────────────┬────────────────┐
         ▼                ▼                ▼                ▼
    Filesystem           Git            Terminal        Processes
  (Jailed Roots)   (Branch Guard)    (Bounded execve)  (Tree Superv.)

Mandatory Invariant: No Filesystem, Git, Terminal, or Process operation may ever execute without traversing the full Authentication $\to$ Policy Engine $\to$ Audit Layer pipeline. In RC-01, this invariant is satisfied via the Minimal Security Kernel, which enforces default-deny admission, canonical workspace root binding, and audit logging before any tool runs.


3. The 20 Permanent Security Principles

  1. Zero implicit trust

  2. Default deny

  3. Least privilege

  4. Fail closed

  5. Separation of authentication and authorization

  6. Mandatory policy mediation for every privileged operation

  7. No policy bypass

  8. Authorized filesystem roots only (strict canonical jailing & descriptor containment)

  9. Secrets and credential stores permanently inaccessible by default

  10. Policy-controlled subprocess execution without shell interpolation

  11. Destructive operations denied by default

  12. Zero cloud-root or production access in public defaults

  13. Universal auditability for every privileged action

  14. Protected Git branches (main, release/*) immutable to agents

  15. Explicit human approval for sensitive mutations

  16. Prohibition of private CES secrets, infrastructure details, and tokens

  17. Permanence of committed secrets (Git history treated as public and forever)

  18. Vendor neutrality across AI model providers and platforms

  19. Code-enforced security invariants (never documentation alone)

  20. Mandatory negative controls proving blocked operations fail


4. Documentation Index

The complete architecture and security foundation of CesSpace ARC is documented below:

Document

Description

Architecture Overview

High-level system architecture, component breakdown, MCP standard, and architectural invariants.

Trust Boundaries

The 4 security zones and 3 trust boundaries isolating untrusted agents.

Threat Model

STRIDE analysis, AI threat vectors (prompt injection, symlink escapes, DoS).

Security Invariants

The 20 inviolable security invariants and their verification criteria.

MCP Tool Taxonomy

Formal schemas and permission tiers for all read-only, mutating, and composite tools.

Permission Model

ALLOW, REQUIRE APPROVAL, DENY outcomes, precedence, Minimal Security Kernel, and approval lifecycle.

Filesystem Boundary Model

Workspace jailing, canonical realpath checks, symlink containment, Linux openat2, and blacklist.

Audit & Evidence Model

Append-only event schemas, data minimization first, redaction, and tiered anchoring.

Structured Error Model

Fail-closed error taxonomy and anti-leakage information sanitization.

Package Ownership

Monorepo package layout, component boundaries, and dependency DAG.

Architecture Decision Records

Formal ADRs capturing key design and security decisions.

RC-01 Scope & Acceptance

Detailed scope, Minimal Security Kernel, tool list, and negative control criteria for RC-01.

RC-02 Scope & Acceptance

Detailed scope, controlled terminal and process execution, and negative control criteria for RC-02.

RC-03 Scope & Acceptance

Detailed scope, safe file modification and patch engine, and negative control criteria for RC-03.

RC-04 Scope & Acceptance

Detailed scope, declarative policy engine, approval state machine, and the 38 frozen negative controls.

RC-04 Final Integration Report

RC-04 acceptance evidence: architecture summary, control coverage table, and final quality gates.

RC-05 Scope & Acceptance

Detailed scope, secure remote gateway, Streamable HTTP over TLS 1.3, mTLS, and 79 negative controls.

RC-05 Final Integration Report

RC-05 acceptance evidence: architecture summary, control coverage table, and final quality gates.

Engineering Governance

Mandatory project engineering, stage-gate, and security governance rules.

Security Policy

Vulnerability disclosure, responsible reporting, and safety invariants.

Contributing Guide

Guidelines for contributing code, tests, and security negative controls.


5. Toolchain & Monorepo Configuration

CesSpace ARC is standardized on:

  • Node.js: ^24.0.0

  • Package Manager: pnpm@12.4.2 (pnpm-workspace.yaml)

  • Internal Dependency Protocol: workspace:*

  • Access: All internal packages are marked private: true during development stages.


6. Development Roadmap & Stages

Stage

Name

Target Capabilities

Status

RC-00

Architecture & Security Foundation

Trust boundaries, threat models, invariants, tool taxonomy, ADRs.

Implemented

RC-01

Read-Only MCP Core

Minimal Security Kernel, 9 read-only tools, canonical jailing, negative controls.

Implemented

RC-02

Controlled Terminal & Processes

Bounded process execution, output limits, timeout enforcement.

Implemented

RC-03

Safe File Modification

Jailed file writing, atomic patches, size limits, approval gating.

Implemented

RC-04

Policy Engine & Approvals

Declarative YAML policy engine, authenticated admin channel, approval tokens, approval audit lifecycle.

Implemented

RC-05

Secure Remote Gateway

Streamable HTTP over TLS 1.3 with SSE response framing, mutual TLS, device enrollment, volatile sessions.

Implemented

RC-06

Audit & Evidence

Append-only JSONL logging, tiered anchoring, automated redaction.

Planned

RC-07

Engineering-Aware Tools

Composite verification commands (arc_verify, arc_stage_evidence).

Planned

RC-08

Integrations & Security Review

Cross-client validation, penetration testing, fuzzing.

Planned


7. License

CesSpace ARC is licensed under the Apache License, Version 2.0.

Related MCP Connectors

Related MCP Servers

  • A
    license
    Not graded
    quality
    D
    maintenance
    Enables secure remote access operations through SSH, SFTP, rsync, VPN, and tunneling with enterprise-grade policy enforcement and audit logging. Provides AI assistants with secure, policy-driven access to remote systems while maintaining comprehensive audit trails and zero-trust security.
    1
    Apache 2.0
  • A
    license
    B
    quality
    A
    maintenance
    Provides policy-driven, auditable SSH access to server fleets for AI assistants with zero-trust security controls, command whitelisting, and comprehensive audit logging to safely manage infrastructure.
    13
    27
    Apache 2.0
  • A
    license
    Not graded
    quality
    A
    maintenance
    Enables AI assistants to securely execute SSH commands on remote servers with connection pooling, session isolation, and a web audit panel.
    3
    MIT
  • F
    license
    Not graded
    quality
    A
    maintenance
    Give AI agents Zero-Trust access to production infrastructure without the risks of granting them shell access. Actions are bounded by policy and an on-host runner.
    335
    -