Two commands that carry most of the layer’s integration weight: tessl mcp start and tessl mcp proxy.
mcp start is how every harness talks to the registry. It is the stdio MCP server that tessl-init writes into each agent’s config, and through it an agent reaches the context in installed plugins as tools rather than as files. That single choice is what makes the product harness-agnostic: MCP is the one extension surface Claude Code, Cursor, Codex, Gemini CLI, Copilot, OpenCode, and OpenClaw all share, so targeting it reaches seven harnesses without seven integrations. The integrates_with: edges on tessl rest on this command.
mcp proxy is the workspace MCP gateway. A workspace declares MCP servers centrally and any agent connects to them through the proxy, so the servers an organization sanctions — and the credentials they need (tessl-environment) — are administered in one place rather than configured per developer. That is the supply-chain argument extended from instructions to tools: if a registry governs which skills may be installed, the same boundary is the natural place to govern which tools an agent may reach.
Worth noting against the wiki’s harness matrix: MCP appears there as a per-harness capability row. This layer treats it as the portability substrate — the thing that makes one capability run everywhere — which is a different use of the same primitive.
See Also
- tessl-init — writes the
mcp startentry into each detected agent. - tessl-environment — the credentials the gateway carries.
- claude-code · opencode — the paged harnesses reached this way.