agent-as-a-service

<div align="center"> # AaaS — Agent as a Service **Turn a working session into an Agent anyone can use.** Publish a Codex or Claude Code session once. Teammates, clients, or their own Agents can continue from what it already learned—without seeing or changing your original conversation. [](https://github.com/Equality-Machine/agent-as-a-service/actions/workflows/ci.yml) [](LICENSE) [](package.json) [](https://aaas-agent-service.b4yesc4t.chatgpt.site) [Try the web app](https://aaas-agent-service.b4yesc4t.chatgpt.site) · [Installation](docs/INSTALL.md) · [Architecture](docs/ARCHITECTURE.md) · [Contributing](CONTRIBUTING.md) · [中文](README.zh-CN.md) </div>  ## Why AaaS The hard part of a useful Agent is rarely the first prompt. It is the context earned over dozens of turns: what matters, what has been tried, what good looks like, and how the work gets done. AaaS makes that head start reusable. | Start with context | Share anywhere | Keep conversations private | | --- | --- | --- | | Publish what a Codex or Claude Code session has already learned. | Give someone a link or Agent ID for the web, Codex, Claude Code, MCP, or HTTP. | Every use starts an independent conversation. Consumer messages never write back to the publisher's source session. | ## A different starting point AaaS is not another prompt builder. It starts from a session where the work has already happened and turns that accumulated understanding into a reusable entry point. | What you share | What the next person receives | | --- | --- | | Prompt or Agent configuration | Instructions for a new conversation | | Shared chat | A readable snapshot of the past | | Workflow export | A definition to import and reconnect | | **AaaS Agent** | **A private continuation of proven working context** | ## One link, one head start 1. A researcher spends a long Codex session testing assumptions and refining a method. 2. They publish it as a **Market Research Agent** and share the link. 3. A product lead opens it and asks the actual next question: > Turn the evidence into a launch brief for enterprise buyers. 4. The product lead gets a private, continuous conversation. The researcher's original session stays unchanged. The same pattern works for project handoffs, reusable coding and research partners, and customer-facing experts: preserve the hard-won understanding, then let the next person start with the outcome. ## Quick start ### Use a shared Agent The fastest path needs no installation: 1. Open the [web app](https://aaas-agent-service.b4yesc4t.chatgpt.site). 2. Paste an Agent ID such as `agt_...`, or open a shared Agent link. 3. Start working. To use shared Agents inside Codex or Claude Code: ```bash npx -y Equality-Machine/agent-as-a-service ``` This installs the AaaS Skill and MCP integration. It does **not** install a Runner or start a background service. Consumers do not need one. After restarting the client, ask naturally: ```text Use Agent agt_... and ask it to turn this idea into a three-step plan. ``` ### Publish your own Agent If this machine will publish sessions, install the publisher role: ```bash npx -y Equality-Machine/agent-as-a-service publisher ``` Then, from the Codex or Claude Code session you want to share: ```text Publish this conversation as an Agent named "Research Partner". Run it on my local machine. ``` A local publisher uses a persistent Runner on the publisher's machine. The Skill explains that system change and asks for confirmation before installing it. Publishing returns an Agent ID and a shareable link. See [Installation and roles](docs/INSTALL.md) for the curl fallback, server Runner setup, supported overrides, and manual configuration. ## How it works ```mermaid flowchart LR S["Private source session"] --> V["Immutable Agent version"] V --> A["Stable Agent ID"] A --> C1["Consumer conversation A"] A --> C2["Consumer conversation B"] C1 --> F1["Provider-native fork A"] C2 --> F2["Provider-native fork B"] ``` - The **source session** stays private on the publisher or paired Runner. - An **Agent version** freezes a completed point in that session. - The **Agent ID** is the stable identity people share. - Every **conversation** receives its own provider-native fork and can continue independently. Local mode keeps source material and provider login on the publisher's machine. Cloud mode transfers an authenticated, AES-256-GCM encrypted source capsule to a paired server Runner. In both modes, the control plane coordinates public metadata, conversations, messages, and job leases. Read the [architecture guide](docs/ARCHITECTURE.md) for the object model, Runner topology, isolation boundary, job lifecycle, and current trust model. ## Interfaces | Interface | Best for | | --- | --- | | [Web](https://aaas-agent-service.b4yesc4t.chatgpt.site) | Using an Agent from any browser | | Skill + MCP | Natural use and publishing from Codex or Claude Code | | HTTP API | Product integrations and automation | | Local Runner | Agents that depend on private files, local repos, or desktop credentials | | Cloud Runner | Always-on execution on a controlled server | The MCP surface includes: ```text runner_status install_local_runner publish_current_agent find_agent agent_start agent_continue agent_end ``` Consumers use only the Agent tools. Runner checks and installation happen only when a user explicitly asks to publish locally. ## Project status AaaS is an **open public preview**. The core fork-safe lifecycle, Codex and Claude Code runtimes, local and cloud Runners, web client, MCP surface, job leases, cancellation, and encrypted cloud capsules are implemented and covered by automated and real-provider acceptance tests. The public control plane does not yet include accounts, quotas, billing, or a complete multi-tenant abuse-prevention layer. Write-enabled tools also require per-conversation workspace, identity, secret, and approval isolation before they are safe for untrusted public traffic. See the [roadmap](docs/ROADMAP.md), [verification evidence](docs/VERIFICATION.md), and [security policy](SECURITY.md) before production adoption. ## Repository guide | Path | Purpose | | --- | --- | | `src/` | Control client, publisher, Runner, runtime, MCP, and local service code | | `cloud/` | Hosted control plane, web app, D1/R2 integration, and cloud tests | | `skills/aaas/` | Installable Codex and Claude Code Skill | | `tests/` | Unit, contract, distribution, and gated real-runtime acceptance | | `docs/` | Architecture, installation, development, roadmap, and verification | Start with the [documentation index](docs/README.md) or the [development guide](docs/DEVELOPMENT.md). ## Contributing Issues, design discussions, documentation fixes, runtime adapters, and test improvements are welcome. Please read: - [Contributing guide](CONTRIBUTING.md) - [Code of conduct](CODE_OF_CONDUCT.md) - [Support guide](SUPPORT.md) - [Governance](GOVERNANCE.md) For vulnerabilities, use GitHub's private security-advisory flow described in [SECURITY.md](SECURITY.md). Never post API keys, Runner tokens, private session files, transcripts, or local paths in a public issue. ## License [MIT](LICENSE) © 2026 Efflora contributors.