OpenCode

What it is

OpenCode is an open-source terminal coding agent whose organising idea is in the spec's model field: any. Frontier APIs, cheaper hosted models, open-weight models behind a gateway, or something running locally on your own hardware — the loop, the tools and the permission rules stay the same and the provider is configuration. Every other terminal agent on this board except OpenClaw ties you to one vendor's models.

That matters more than it sounds. Model quality and pricing have moved repeatedly over the past two years. A harness you can re-point in a config file turns each of those moves into a small decision instead of a migration.

How the loop works

Read, search, edit, run, check, with shell access and MCP for anything external. Project instructions come from a Markdown file in the repository, following the same convention Codex CLI uses, so standing instructions move between the two without rewriting. Language-server integration gives it real symbol information rather than text search alone, which is the quiet reason its edits land more often than a grep-driven agent's.

It is built as a client and a server rather than one process. The agent runs somewhere — your machine, a remote box — and clients attach to it. In practice that means a session can outlive the terminal you started it in, and you can drive the same session from more than one place.

Isolation and permissions

Permissioned tools, configured per project: which commands may run unattended, which require a prompt, which are refused. There is no OS-level sandbox in the box, so the boundary is the policy you write plus wherever you chose to run it. Container or worktree, as ever, if the work is unfamiliar or the repository is not yours.

Provider-agnosticism adds a consideration the single-vendor agents do not have: every provider you configure is another party seeing your code, with its own retention and training terms. A gateway is one more hop again. That is the trade for choice — worth making deliberately, and worth writing down for anyone else on the team.

Who it is for

People who want the model to be a variable: cost-sensitive teams, anyone evaluating models seriously, polyglot shops, and anyone who needs a local model for some work and a frontier model for the rest. Also anyone who simply refuses to make a proprietary binary load-bearing in their development workflow.

The project runs its own subscription gateway alongside the agent, bundling access to a set of open-weight coding models behind one key that works with any client speaking the common APIs. Our write-up of the Go plan covers the arithmetic and the catch: the caps are dollar-denominated over a rolling window, so context-heavy agent sessions hit the short limit long before the monthly allowance runs out. Check the current figures in their docs rather than trusting any page's snapshot.

Limits

• Agnostic means uneven. Prompts and tool descriptions are tuned against some models and merely work with others. A weaker model in the same harness produces visibly worse sessions.
• No sandbox by default. Permissions are rules, not confinement. Codex CLI is the stronger choice if isolation is non-negotiable.
• Configuration surface. Model choice is a lever, and levers need someone to set them. Teams that want one sanctioned path may find the flexibility a cost.
• Terminal only. No first-party editor experience; pair it with an editor rather than expecting one.
• Community-maintained. Fast-moving, no support contract, and the gateway's commercial terms can change independently of the open-source client.

Alternatives on this board

• OpenClaw — the other open, model-agnostic terminal agent, broader in scope and less repository-focused.
• Codex CLI — open source with a default sandbox, at the cost of model choice.
• Claude Code — the most polished loop, if a single vendor is acceptable.
• Gemini CLI — Apache-2.0 and unusually strong on whole-repository context.