OPENCODE GOES PROTOCOL-FIRST: MCP TOOLS, LSP CODE INTEL, AND A PLAN/BUILD AGENT SPLIT
OpenCode now treats coding agents like IDEs: protocol-first, model-agnostic, with MCP tools, LSP code intel, and a Plan/Build split. This [explainer](https://w...
OpenCode now treats coding agents like IDEs: protocol-first, model-agnostic, with MCP tools, LSP code intel, and a Plan/Build split.
This explainer breaks down how OpenCode separates the agent runtime from the model and runs as TUI, desktop, or IDE extension with BYOK, the Zen gateway, or the Go plan.
Under the hood, sessions persist locally in SQLite; code intel comes from LSP; external tools attach via MCP; and a read-only Plan agent gates the write-capable Build agent for safer changes details.
The trade-off is clear: protocol-level extensibility and provider freedom versus setup overhead for MCP servers, LSPs, and permissions more.
Protocol-first design reduces vendor lock-in and lets you standardize tools across models.
Plan-before-Build adds a review gate that can cut risky writes in large repos.
-
terminal
Run Plan-only on a service repo, then authorize Build; compare diffs and failure rates against your current agent or IDE workflow.
-
terminal
Register one MCP server and switch models via BYOK vs the Zen gateway; measure tool-call latency, token cost, and permission prompts.
Legacy codebase integration strategies...
- 01.
Reuse existing LSP servers from your stack and gate file/shell tools with least-privilege; centralize BYOK key management.
- 02.
Pilot the TUI in CI sandboxes to validate Plan/Build flows before enabling writes in monorepos.
Fresh architecture paradigms...
- 01.
Design agent workflows around MCP from day one; expose internal services as MCP servers with auditable permissions.
- 02.
Pick default models per language and codify approval rules between Plan and Build.
Get daily OPENCODE + SDLC updates.
- Practical tactics you can ship tomorrow
- Tooling, workflows, and architecture notes
- One short email each weekday