Cloud Harness MCP gives ChatGPT, Claude, Cursor, and Codex full execution power over isolated repository clones inside bounded Docker executors — plus managed coding agents, one MCP gateway for every other server, and a shared skill library. Docker host sockets and GitHub App secrets never leave the control plane.
One credential-free ingress, a trusted control plane that owns every secret and the Docker socket, and disposable execution containers that run repository code, managed agents, and Git transfers on separate rails.
Cloudflare Access Managed OAuth is the authorization server at the perimeter, and the API verifies only the signed assertion Access forwards. Dashboard API keys use a separate lane. The ingress proxy holds no secrets and never joins the control network. How the OAuth lane works.
The API speaks MCP, serves the dashboard, fronts your downstream MCP servers, and delivers signed event webhooks. The runner alone owns the Docker socket, encrypted state, and GitHub App credentials, and feeds a separate model gateway that keeps provider keys out of every container.
Repository code runs in a non-root executor with a read-only root filesystem and resource caps. Dependency egress is the shipped default and network-none is the per-workspace or instance-wide opt-out. Managed agents and Git transfers each get their own short-lived container.
A high-velocity, deterministic sequence from workspace boot to safe git push and automatic sandbox purging.
agent@claude-code:~$ call mcp.workspace_open({
"repositoryUrl": "https://github.com/org/core-infra",
"idempotencyKey": "task_448f_09a1"
})
[200 OK] Streamable HTTP Connection Established (14ms)
├── Edge: Cloudflare Access accepted OAuth bearer; origin verified signed assertion
├── Runner: Ephemeral GitHub broker minted stdin-only push key
├── Docker: Spawned non-root executor (TTL: 1800s, net: dependency-access)
└── Workspace Ready: ws_8f93e120da4c71b2e9f0
agent@claude-code:~$ call mcp.exec_run({
"workspaceId": "ws_8f93e120da4c71b2e9f0",
"command": "npm test -- --bail"
})
✓ Test suite completed in 1.42s (0 host leaks, 0 socket mounts)
Register GitHub, Supabase, PostHog, Cloudflare, or any remote MCP server once. Your clients connect to a single /mcp-gateway URL whose tool list never grows, so adding servers never eats your context window.
searchinspectexecutepermissionsstatusThe gateway always exposes the same five tools. Agents search the cached catalog, inspect one tool's schema, then execute it, instead of loading hundreds of tool definitions up front.
Header values can point at a write-only global secret, and OAuth-protected servers connect from the dashboard. Clients never see the downstream credential.
Set a default allow or deny per server, override single tools, and review every call in the dashboard. Outbound sockets are validated and address-pinned.
Your client delegates a bounded task with agent_spawn, keeps working, and collects the result later. Each agent runs in its own disposable container against a workspace, with models reached only through the trusted model gateway.
A bounded Pi coding session with token, cost, and time budgets, ten safe file and search operations, and live status, logs, and follow-up messages.
Run Anthropic's coding agent as a managed runtime under the same container, lease, and budget boundary.
Run OpenAI's coding agent as a managed runtime under the same container, lease, and budget boundary.
Store keys for OpenAI, Anthropic, OpenRouter, Google Gemini, DeepSeek, OpenCode Go, or any HTTPS endpoint you trust, then build model profiles with pricing, budgets, and tool permissions. Draft and validate profiles from your client with the model_* tools; keys, activation, and spend policies stay dashboard-only. No restart needed.
Provider keys are write-only and encrypted at rest. Agents get an opaque, short-lived lease instead, and in-flight runs stay pinned to the exact profile revision they started with.
Every request reserves its worst case before it goes upstream. Set daily or monthly spend policies per owner or per profile, and read the totals with model_spend_summary.
agent@chatgpt:~$ call mcp.agent_spawn({ "workspaceId": "ws_8f93e120da4c71b2e9f0", "profileId": "pi-default", "idempotencyKey": "fix_auth_01", "prompt": "Fix the flaky auth test" }) → accepted: agent_41c2e9b07d1f4a8c3b6e (generation 1) agent@chatgpt:~$ call mcp.agent_status({ "workspaceId": "ws_8f93e120da4c71b2e9f0", "agentId": "agent_41c2e9b07d1f4a8c3b6e" }) ✓ completed • summary ready • 41% of cost budget used
Skills are folders of instructions and optional scripts. Manage them once in the dashboard and every connected client discovers them in each workspace through skills_list, skills_read, and skills_run.
Every revision is append-only. Diff, restore, fork, and archive skills, and see which skill sets or live workspaces pin each one.
Group skills into sets and let skill_suggest recommend the matching skill before an agent plans its work.
Mount curated toolkits such as mattpocock/skills and obra/superpowers, or any pinned public Git repository of skills.
Licensed AgentKit.best kits such as engineer and marketing mount read-only from a signed registry, and the ak CLI ships inside the executor. The licence token never reaches repository code.
AgentKit.best packages development and marketing workflows as skills, hooks, and specialist subagents, with the ak CLI and a desktop app to install and manage them. Cloud Harness MCP is built by the same team, so licensed kits mount read-only in every workspace and ak is already inside the executor.
From file editing and symbol search to Git, GitHub, packages, knowledge, and managed coding agents — every tool typed, bounded, and owner-scoped.
Search LSP symbols, declarations, and definitions across TypeScript, Go, Python, and Rust.
Apply line-anchored atomic text patches with conflict detection and rollback.
Rust-powered regex search with cross-line multiline support and gitignore respect.
Interactive persistent PTY shell sessions supporting full terminal interaction and ANSI signals.
Dispatch DAG multi-stage tasks with topological barriers, artifact tracking, and timeouts.
Spawn instant isolated branch worktrees allowing subagents to work without Git collisions.
Push commits through stdin-fed short-lived GitHub App tokens with force-with-lease safety.
Discover and execute repository-defined custom skills, automated code reviewers, and linters.
Run repository lifecycle hooks (pre-commit, post-build, security-scan) in isolated sandboxes.
Trigger owner-reviewed repository deployment recipes directly from AI agent context.
Store and recall durable repository decisions and conventions across agent sessions.
Spawn a budgeted Pi coding agent in its own isolated container against a workspace.
Read an agent's state, budgets, bounded logs, and terminal summary without blocking your session.
Create and validate model profiles, verify provider connections, and review spend from your client.
Connect the AI tools you already use — ChatGPT, Claude, Claude Code, Cursor, Codex, Gemini, Grok, Antigravity — to one remote harness. Plan in a chat, then clone, build, test, review, and open the pull request without leaving it.
Honest status: it is not a fully autonomous software factory yet — but with managed agents, the gateway, and shared skills, we are getting very close.
The owner deployment has two non-interchangeable lanes: Managed OAuth at https://harness.zuey.me/mcp, where Cloudflare Access is the OAuth authorization server, and a static API-key lane at https://api.harness.zuey.me/mcp. How the OAuth lane works.
Static API-key lane: connect via Streamable HTTP with a dashboard-managed API key. Claude Code will instantly acquire all remote execution capabilities.
claude mcp add cloud-harness \ --transport http \ https://api.harness.zuey.me/mcp \ --header "Authorization: Bearer YOUR_API_KEY"
Deploy a private single-owner server on your own Linux VPS, install the portable companion skill, or integrate agent plugins.
Single-command interactive installer for Ubuntu 24.04 / Debian 12 with automatic Let's Encrypt TLS reverse proxy to loopback 127.0.0.1:3100.
cloudharness CLI (status, token view/rotate, upgrade)Runs repository installer as root. Review source on GitHub repository first.
curl -fsSL https://raw.githubusercontent.com/bestagentkits/cloud-harness-mcp/main/scripts/install.sh | sudo bash
Designed intentionally for one trusted operator or team. Arbitrary execution capability inside shared-kernel executors with strict boundary separation.
The API and Ingress proxy never receive Docker socket mounts or host volume binds. Docker daemon control remains strictly localized to the trusted Runner daemon.
GitHub App private keys stay inside the runner. Installation tokens are short-lived and piped over stdin to an ephemeral git helper.
Executors run as non-root UID 10001 with read-only root filesystems and memory/CPU caps. Dependency egress is the shipped default and is never a silent downgrade; choose network-none per workspace or instance-wide for full isolation.
Workspaces are destroyed immediately upon workspace_close or when TTL expires, leaving zero leftover state, containers, or leaked temp files.
No subscription, no seat limits, no vendor lock-in. Run it on your own VPS, read every line, and help shape where it goes next — issues, pull requests, and new skills are all welcome.