Grist

Grist is an open-source coding-agent harness that wraps OpenCode v2 with model routing, engineering guidance, spend controls, and specialist workers. It is aimed at developers and software teams that want an agent to choose an appropriate model for each task instead of sending every request to the same expensive option. A Jev confidence gate evaluates the work, assigns it to one of four cost tiers, and routes the request to the model configured for that tier. The coding loop remains based on OpenCode, while Grist concentrates on the operational layer around it. That includes how a model gets chosen, what team rules enter the context, when a run can move upward to a more capable tier, and how usage is reported.

The project ships its routing behavior as the @grist-ai/plugin package inside the Grist binary. Every session loads the confidence gate, doctrine injector, and control-plane hooks without a separate plugin setup. The gate can select a cheaper, medium, frontier, or premium rung according to its confidence in the task. Model switching during a run is escalate-only, so the harness can move to a stronger tier when needed without dropping to a cheaper one partway through. An effort setting also adjusts how much exploration happens within the selected rung. Straightforward tasks get a tighter budget, while ambiguous work gets more room to investigate. Users can override that behavior for testing, but the intended workflow lets the gate decide.

Doctrine gives the agent persistent engineering standards for the session. That can keep repository expectations and working conventions present without repeating the same instructions in every prompt. Grist also provides named specialist subagents for exploration, planning, review, and verification. They run within the rung already selected for the parent task, which prevents decomposition from silently promoting the work to a more expensive tier. Repeated specialist calls can resume their previous sessions so context carries forward instead of starting cold each time. Together, those parts target a common agent problem. The code-writing loop may be capable, but a team still needs consistent context, task decomposition, verification, and an explicit policy for when extra model capability is worth the cost.

Grist retains the broader OpenCode v2 command surface for authentication, model inspection, debugging, MCP, sessions, pairing, upgrades, and statistics. The serve command exposes an API and web interface, and a managed service mode keeps one background server available while the run command acts as a thin client. Developers can run a task from the terminal inside a repository and consume event output as JSON for scripts. Sessions can be resumed by identifier, and a published Grist skill teaches another personal AI agent to delegate multi-step coding jobs to the CLI. The project also has a doctor command that validates gateway credentials with an authenticated request, which is more useful than checking only whether a key exists.

The economic model is bring your own key. Inference is billed to the user's provider account through OpenRouter, Vercel AI Gateway, or another OpenAI-compatible endpoint, and the project states that it adds no markup and never holds the user's inference budget. Its gateway maps rung names to actual models, enforces hard spending caps, and stores per-user provider keys with AES-256-GCM encryption without logging them. Usage metering prefers the provider's reported cost and accounts for cached tokens at their cached rates. The dashboard can therefore reflect what the underlying provider charged rather than an independent estimate. Teams that need full control can also use the repository's gateway components and Docker packaging as the basis for their own deployment.

Grist installs globally from npm and is licensed under MIT, so the harness itself can be inspected, modified, and used without a software subscription. Model inference is still metered by the chosen provider, which means open-source pricing doesn't make runs cost-free. It best suits developers comfortable with command-line agent workflows who want automatic cost-tier routing and consistent engineering doctrine. It is a young project with a small public community, and its own benchmark figures should be read as internal evidence rather than independent validation. The product site was not reachable through the research browser, but the active official repository provides installation, architecture, pricing posture, license, and current feature details. Buyers should also avoid confusing this coding harness with the older relational spreadsheet product that shares the Grist name.