MCP Architecture & OAuth
Relevant source files
The following files were used as context for generating this wiki page:
The Model Context Protocol (MCP) in the Oh-My-OpenCode project provides a standardized interface for agents to interact with external tools, documentation, and web resources. It operates as a three-tier system comprising built-in services, Claude Code compatible configurations, and skill-embedded definitions. This architecture enables agents like Sisyphus and Atlas to maintain a lean context window by offloading specialized data retrieval to external MCP servers.
The system also includes specific notices regarding OAuth. As of early 2026, the project explicitly states that it does not provide custom implementations of restricted third-party OAuth signatures. Instead, it relies on standard integration methods while warning users about the security implications of unofficial third-party impersonation sites.
Sources: AGENTS.md:214-219, README.md:107-124
MCP System Tiers
The MCP architecture is organized into three distinct layers to maximize compatibility and extensibility.
| Tier | Description | Key Components |
|---|---|---|
| Built-in | Native project MCPs for essential research. | Exa (Web Search), Context7 (Docs), Grep.app (GitHub Search) |
| Claude Code Compat | Support for .mcp.json with variable expansion. | ${VAR} support for environment-specific configs |
| Skill-embedded | YAML frontmatter definitions within skills. | Integrated tool definitions in src/features/builtin-skills/ |
Sources: AGENTS.md:214-219, README.md:183-186
Built-in MCP Services
The project prioritizes three specific MCP servers for automated research during the planning and execution phases. These are primarily utilized by the Librarian and Explore agents to gather external context.
- Exa: Used for broad web searches to find library best practices or troubleshooting info.
- Context7: Specialized in retrieving and parsing official technical documentation.
- Grep.app: Facilitates searching public GitHub repositories for real-world implementation examples.
Sources: README.md:183-186, src/agents/prometheus/interview-mode.ts:211-218
OAuth and External Authentication
The project maintains a strict stance on OAuth security and official access. A dedicated notice clarifies the handling of Anthropic's OAuth restrictions.
Third-Party OAuth Access
The documentation warns that third-party OAuth access has been restricted by providers. The project explicitly disclaims responsibility for unofficial tools that spoof request signatures and confirms it does not include custom implementations of restricted OAuth systems.
Sources: README.md:107-124
Security Warnings
Users are cautioned against impersonation sites (e.g., ohmyopencode.com) which are not affiliated with the project. Official distributions are strictly limited to the GitHub repository releases.
Sources: README.md:1-12
MCP-Driven Agent Workflow
MCPs are integrated into the lifecycle of specialized agents. The flow typically moves from high-level planning to MCP-assisted research.
The diagram shows how MCP-enabled agents (Librarian/Explore) feed data into the planning pipeline.
Sources: src/agents/prometheus/plan-generation.ts:47-66, src/agents/metis.ts:89-105
Librarian Agent and MCP Integration
The Librarian agent is the primary consumer of MCP services. It uses GitHub CLI and Web Search (via MCP) to perform multi-repository analysis and retrieve implementation examples.
The sequence illustrates the Librarian agent leveraging MCP for external data synthesis.
Sources: AGENTS.md:204-211, sisyphus-prompt.md:224-237
Data Flow in Planning and Execution
During the plan generation phase, the Metis agent uses MCP-derived findings to identify gaps. Once a plan is approved by Momus, the Atlas orchestrator manages the execution, potentially re-invoking MCPs if subagents encounter unfamiliar libraries.
| Agent | MCP/Tool Interaction | Purpose |
|---|---|---|
| Librarian | Exa, Context7, Web Search | External documentation and OSS examples. |
| Explore | Grep.app, Local Grep | Internal and external codebase pattern discovery. |
| Metis | Reviewing Research Findings | Catching unstated requirements or scope creep. |
| Atlas | delegate_task to Librarian | Resolving issues with unfamiliar packages during implementation. |
Sources: src/agents/metis.ts:162-175, src/agents/atlas/default.ts:258-261, src/agents/momus.ts:114-124
Plan Generation Protocol
When the planner (Prometheus) generates a work plan, it must incorporate "Inherited Wisdom" from background research. This wisdom is stored in a notepad structure within .sisyphus/notepads/ and consists of MCP search results.
Sources: src/agents/atlas/default.ts:109-119, src/agents/prometheus/plan-generation.ts:80-92
Summary of Integration Points
The MCP system is not just a toolset but a core part of the agent orchestration logic. It allows the Sisyphus and Atlas agents to operate as "AI Managers" who delegate the heavy lifting of information gathering to specialized subagents backed by MCP protocols. This ensures technical accuracy by grounding agent actions in real-world documentation and existing code patterns rather than model hallucinations.
Sources: README.md:154-162, src/agents/atlas/default.ts:17-25