OpenCode 2.0: one daemon behind every surface, and every MCP server becomes a function the agent scripts
OpenCode 2.0 went to Hacker News on October 14 and sat at 848 points with 370 comments, two days after the project's own launch post passed 719,000 impressions. The repo is at 214,000 stars. The dating is messy in a way worth recording: the v2.0.0 git tag is from September 11, the npm package has shipped 2.0.x since then, the announcement note is dated October 12, and the GitHub Releases page still lists v1.19.3 from October 16 because the 2.x line is tagged but never got Release objects. By October 16 the current tag was v2.0.31. Last week we cut this story as unverifiable on the Releases page. The tags were the evidence.
The architecture change is the headline. In 2.0 the server that used to be a power-user option now runs as a single background process on your machine, and the TUI, the web interface and the desktop app all talk to that one process. A new client package autodiscovers it, which the team pitches for scheduled tasks, and a pair command exposes it to other hosts, with a confirmation prompt before it opens the port. The server API was rewritten from scratch. The note's own verdict on the old one: it grew haphazardly as we and the world collectively learned what agents are and should be. It was bad. OpenCode now calls its own API, for example to move a session into a git worktree when new work starts. One user shipped a plugin within a day that tunnels the running daemon through Cloudflare Access so you can keep prompting from a phone after closing the terminal.
Code Mode is the design bet. A minimal JavaScript interpreter turns MCP servers and OpenAPI endpoints into functions the agent calls from a script: call a tool, map over the results, call the next tool, one round trip instead of one call per step. Every MCP server is opted into Code Mode by default, and the claim is you can add a thousand of them with minimal context bloat because the agent discovers and composes them in code. That is the same pattern as context-mode, which was back on GitHub trending this week with 30,000 stars and the same pitch of keeping raw tool output out of the window. The counterargument showed up immediately: a script can fire a dozen actions before anyone looks, so safety checks now have to live inside each tool. Code Mode's fetch blocks file access and removing a worktree needs edit approval, which is OpenCode's partial answer. Background subagents, async shell commands you can push to the background with a hotkey, 80 plus built-in plugins with hot reload so the agent can write a plugin mid-session, and an SDK for running your own OpenCode flavor on a server round out the list. Breaking: 1.0 plugins do not work, the old server API is gone, upgrade is opencode upgrade --major.
Two side notes from the same 48 hours. OpenCode is where Reflection's Beam and a stealth model called Cat Jam showed up free this week, which tells you harnesses are now a distribution channel for model vendors. And a Neuralwatt post on October 16 measured 30 days of inference traffic and found every prompt-cache break in its tests came from a harness editing what it had already sent, with a miss roughly doubling a request's energy. The daemon-plus-script shape is now the mainstream open design, with OpenClaw's gateway and Hermes Agent converging on it the same month. What the daemon is allowed to do is the open question.
Links: https://anoma.ly/notes/opencode-2-0/ , https://github.com/anomalyco/opencode , https://opencode.ai/v2/docs/migrate-v1/ , https://news.ycombinator.com/item?id=50094994 , https://x.com/opencode/status/2109649032931119362 , https://neuralwatt.com/news/why-harnesses-matter-for-energy-efficiency
← Back to all articles
The architecture change is the headline. In 2.0 the server that used to be a power-user option now runs as a single background process on your machine, and the TUI, the web interface and the desktop app all talk to that one process. A new client package autodiscovers it, which the team pitches for scheduled tasks, and a pair command exposes it to other hosts, with a confirmation prompt before it opens the port. The server API was rewritten from scratch. The note's own verdict on the old one: it grew haphazardly as we and the world collectively learned what agents are and should be. It was bad. OpenCode now calls its own API, for example to move a session into a git worktree when new work starts. One user shipped a plugin within a day that tunnels the running daemon through Cloudflare Access so you can keep prompting from a phone after closing the terminal.
Code Mode is the design bet. A minimal JavaScript interpreter turns MCP servers and OpenAPI endpoints into functions the agent calls from a script: call a tool, map over the results, call the next tool, one round trip instead of one call per step. Every MCP server is opted into Code Mode by default, and the claim is you can add a thousand of them with minimal context bloat because the agent discovers and composes them in code. That is the same pattern as context-mode, which was back on GitHub trending this week with 30,000 stars and the same pitch of keeping raw tool output out of the window. The counterargument showed up immediately: a script can fire a dozen actions before anyone looks, so safety checks now have to live inside each tool. Code Mode's fetch blocks file access and removing a worktree needs edit approval, which is OpenCode's partial answer. Background subagents, async shell commands you can push to the background with a hotkey, 80 plus built-in plugins with hot reload so the agent can write a plugin mid-session, and an SDK for running your own OpenCode flavor on a server round out the list. Breaking: 1.0 plugins do not work, the old server API is gone, upgrade is opencode upgrade --major.
Two side notes from the same 48 hours. OpenCode is where Reflection's Beam and a stealth model called Cat Jam showed up free this week, which tells you harnesses are now a distribution channel for model vendors. And a Neuralwatt post on October 16 measured 30 days of inference traffic and found every prompt-cache break in its tests came from a harness editing what it had already sent, with a miss roughly doubling a request's energy. The daemon-plus-script shape is now the mainstream open design, with OpenClaw's gateway and Hermes Agent converging on it the same month. What the daemon is allowed to do is the open question.
Links: https://anoma.ly/notes/opencode-2-0/ , https://github.com/anomalyco/opencode , https://opencode.ai/v2/docs/migrate-v1/ , https://news.ycombinator.com/item?id=50094994 , https://x.com/opencode/status/2109649032931119362 , https://neuralwatt.com/news/why-harnesses-matter-for-energy-efficiency
Comments