Agents

Set up your agents

ThreadGrid doesn't sell AI access. It runs each agent's own command-line tool on your computer, signed in as you. Install at least one, sign in once in a terminal, and ThreadGrid finds it.

How ThreadGrid finds an agent

ThreadGrid looks for claude, codex and opencode on your PATH and in the usual install folders (Homebrew, npm, Bun, and each tool's own installer; on Windows also WinGet and Scoop). Then it checks whether you're signed in.

See every agent's state on the Grid page in the sidebar: Ready, Not signed in or Not installed. Just signed in? Click Refresh. Capacity readings refresh every five minutes.

Your provider logins stay inside each tool. ThreadGrid never reads, copies or uploads them.

Claude

Uses the claude CLI, signed in with your Claude plan (Pro or Max) or however you normally sign in.

Install the claude CLI: Anthropic's setup guide. On Windows, install Git for Windows first.
Sign in:

claude auth login

Check it with claude auth status. ThreadGrid reads the same status.

Capacity: observed. ThreadGrid reads the usage Claude reports while it works, so the reading appears after Claude's first run.

Codex

Uses OpenAI's codex CLI, signed in with your ChatGPT plan or an API key.

Install it: github.com/openai/codex.
Sign in:

codex login

Check it with codex login status.

Capacity: exact. Codex reports its own limits and reset times.

Codex runs inside its own sandbox and asks only when it wants to step outside it. The stricter access levels switch it to asking more often.

OpenCode

Uses the open-source opencode CLI. Its free models work without an account. To use your own providers, connect them to OpenCode first.

Install it: opencode.ai/docs.
Optional: connect a provider:

opencode auth login

With no provider connected, ThreadGrid shows OpenCode as Free models. Pick a model with the model chip in the composer after you choose OpenCode.

Capacity: unknown. OpenCode doesn't report plan limits, so ThreadGrid doesn't guess.

API keys

You can add your own Anthropic or OpenAI API key as a fallback for when a subscription runs out. API usage is billed per token to your Anthropic or OpenAI account, so it may cost money.

Open the Grid page. Under your agents you'll see Claude (API key) and Codex (API key).
Paste your key and click Save key. You'll see Key saved.
Choose what happens when the subscription runs out: Only when picked (default) or Use the key.

With Only when picked, the key is used only when you pin the API-key agent in the composer. With Use the key, a task runs on the key once no free agent can take it, and the run shows that it used your key.

The key is encrypted in your Keychain on a Mac, or Windows' credential store on Windows. ThreadGrid never shows, logs or uploads it, and it's never sent to a paired computer or phone. Manage it on the computer you saved it on. The key runs through the Claude Code or Codex CLI, so install that CLI too.

Not in the beta yet. OpenRouter keys and cost budgets aren't supported yet. OpenCode can use a key you give it in its own CLI, and ThreadGrid will drive it as usual.

Choosing and pinning agents

The agent chip in the composer decides who does the work:

Auto routing (default): balances fit for the task against capacity left.
Smart routing: favors the strongest agent for the task.
Saver routing: preserves scarce capacity.
Always use: pin one agent, and optionally a model. A pinned agent is always used, even if Capacity Protection would skip it.

Capacity Protection

On the Grid page, each agent with a usage reading has three settings:

Preserve the last (default 15%): below this share left, routine work goes to other agents.
While preserved, still use it for (default Planning and Review).
When every agent is preserved or out: Spend reserve (default) or Queue until the earliest reset.

Agent not showing up? See Troubleshooting.