OpenCode vs Aider

OpenCode calls itself the open source AI coding agent, and the word to weigh is agent. It is a TypeScript program under the MIT licence from Anomaly; the repository, once under the SST organisation, now lives at anomalyco/opencode. Give it a task and it decides what to read, edits files, runs commands, watches the results, and goes again, under permission rules you set per tool. The architecture is client/server: a headless server (opencode serve, an HTTP API on localhost) with the terminal UI as one client and a web client, a desktop app in beta, and a VS Code-family extension as the others. It ships a build agent with every tool enabled and a plan agent that asks before edits and shell commands, and it moves fast — v1.18.27 landed on 2 September 2026.

Aider calls itself AI pair programming in your terminal, and the word to weigh is pair. It is a Python program under Apache-2.0 from Paul Gauthier, installed with aider-install on Python 3.8 to 3.13. You add files to the chat, say what you want, and Aider edits those files and commits the result — every change a git commit with a generated message, undone with /undo. It does not wander: it lints after edits and runs your tests if you give it a --test-cmd, but any other shell command is one you run yourself with /run. Its pace has slowed: the newest release on PyPI is 0.86.2 from February 2026, three point releases since August 2025, though pull requests were still being merged in May 2026.

An agent owns the loop; a pair programmer shares it. With OpenCode you describe the outcome and watch a transcript of tool calls: the model chooses what to open, what to search, what to run, and how many times to try, and your control is the permission table — ask, allow, or deny per tool, down to a glob. With Aider you own the context: the repo map (a graph-ranked summary of the codebase, about a thousand tokens by default) tells the model what exists, /add tells it what it may change, and each reply is a set of edits you read as a diff before the next turn.

The editing mechanics follow. Aider asks the model for edits in a format chosen per model — whole files, diff search/replace blocks, udiff, or an architect/editor pair where one model proposes and a second writes the edits — and turns every accepted edit into a commit. OpenCode edits through tools, keeps its own /undo and /redo of messages and file changes, and leaves git to you; nothing is committed until you commit it. One approach puts the model’s work in the history where you can audit it; the other keeps the history yours. Which you prefer says more about how you review than about the tools.

Both tools end the same way: a process on a computer, holding a checkout, a login, and hours of work. OpenCode’s server listens on localhost; Aider is a chat in a terminal. Reaching that computer from the other laptop, or from a phone when the agent asks a question at nine, is the same problem for both, and it is the one spawnd is built for. One daemon on each host you own dials out, so nothing listens and no port opens; every session lives on the host, its PTY owned by a worker process, so it survives the closed tab, the dropped connection, and the laptop lid, scrollback intact; and any browser is the console, a phone included. OpenCode and Aider are built-in shortcuts beside Claude Code and Codex — an agent is a named command in a real shell, so any CLI can be added — and they authenticate on the host exactly as they do today; spawnd holds no provider credentials and adds no API-key markup. Pick either; the agents roundup puts both beside Claude Code, Codex, and the rest.