본문으로 건너뛰기

Set the context input limit

The context input limit caps how much material the agent sends to the model in one request. Without it, the agent budgets input automatically, and the ring beside the model picker shows how full the context is. See how full the context is

The context budget lives in the workspace's config.toml. To change this budget, ask the agent or edit that file yourself. Global hooks have their own user-wide ~/.grida/config.toml source and are inspected in Settings → Hooks. Configure hooks · Desktop settings

Ask the agent​

On macOS and Linux, ask in an Act message, for example “cap the context input at 64000 tokens for future messages.” The agent reads and edits its settings file like any other file, and every change to it asks for your approval in the chat, even though the agent edits project files without asking. Allow once covers that one change; the next change asks again. Decline leaves the file as it was.

The agent saves only valid settings. If the new content is invalid, it reports why and the file stays unchanged. A saved change applies to messages you send from then on; work that is already running, paused or queued keeps the value it started with.

Edit the file yourself​

  1. Quit Desktop.

  2. In Desktop's application data folder, open chats/workspaces/. Each workspace has its own folder there, named by a digest of the workspace rather than its path. Desktop does not show which folder belongs to which workspace, so pick the one used most recently by the workspace you mean.

    PlatformApplication data folder
    macOS~/Library/Application Support/Grida
    Linux~/.config/Grida, or Grida under $XDG_CONFIG_HOME if set

    Grida Insiders uses a Grida Insiders folder in the same place.

  3. Create or edit config.toml in that folder, for example:

    schema_version = 1

    [context]
    max_input_tokens = 64000
  4. Start Desktop and open the workspace. The new limit applies to messages you send from then on.

A hard limit needs a connection that can count its requests exactly. If it cannot, or a request is larger than the limit, the turn fails before the model starts answering. Remove max_input_tokens to return to automatic budgeting. An invalid file prevents the workspace's chat from starting, and the error names the problem.

Change it from a client​

Source-built headless runs read the same file from the state directory you choose, and the agent can edit it under a permission profile that allows file writes. ACP clients can also inspect, change and reload the limit. Agent configuration reference