Claude Code

How LinkCode runs Claude Code, including in-app sign-in, the five approval policies, model and reasoning-effort switching, slash commands, and shared thread history.

Claude Code is the most complete integration of the five supported agents: five approval policies, a curated model list, live reasoning-effort control, subagent transcripts, and full thread history shared with the claude CLI.

Set it up

LinkCode doesn't ship the Claude Code CLI. The first time you pick it, LinkCode either finds an install already on your machine or offers to Download the version this release was tested against. See agents for how that resolution works.

Three ways to authenticate, in the order most people use them:

Sign in from inside LinkCode. The Sign in to Claude Code card opens a Claude sign-in page in your browser; if it shows an authorization code rather than completing on its own, paste the code back into LinkCode.
Reuse an existing claude CLI login. If you've already signed in on this machine, LinkCode uses those credentials as-is.
Connect an account under Settings → Providers — a Claude subscription, an Anthropic API key, or a gateway/custom endpoint. See settings.

Approval policy

Five policies, in the order they appear in the composer's shield menu:

Policy	What it does
Ask permissions	Always asks before editing files and running commands.
Accept edits	Auto-approves file edits; still asks for everything else.
Plan mode	Read-only research; proposes a plan before making changes.
Auto mode	A classifier approves routine actions and blocks risky or external ones.
Bypass permissions	Skips all permission prompts in this workspace.

You can switch policies at any point in a thread, and the change applies immediately. Not every policy is available on every account — if a switch is rejected (Auto mode or Bypass permissions on an account without the entitlement), LinkCode keeps whatever policy was active before.

When you haven't picked a policy, Claude Code falls back to whatever permissions.defaultMode is set to in the closest settings file it finds, checked in this order:

.claude/settings.local.json (project, local to your machine)
.claude/settings.json (project)
~/.claude/settings.json (your user settings)

This is the same precedence the standalone claude CLI uses. LinkCode only reads these files — it never writes to them.

For the difference between an approval policy (a standing setting) and a one-off permission prompt, see approvals.

Model and reasoning effort

Claude Code has a curated model list:

Fable 5
Opus 5
Opus 4.8
Opus 4.7 (Legacy)
Opus 4.6 (Legacy)
Sonnet 5
Sonnet 4.6 (Legacy)
Haiku 4.5

Reasoning effort runs Low, Medium, High, xHigh, Max, and Ultracode. Model and effort both switch live, mid-thread, with one exception: Max can't be applied to an already-running process, so LinkCode restarts the underlying Claude Code process and resumes your thread in place. There's a brief pause, but no lost context. The same restart happens when switching away from Max.

Ultracode additionally requires dynamic workflows to be enabled in your Claude configuration; if it isn't, the switch is rejected and your previous effort level stays selected.

Commands

Start a message with / to run one of Claude Code's own commands — LinkCode lists what your setup actually advertises, including anything you've defined yourself. /usage is handled specially: LinkCode reports your usage in-thread instead of forwarding the command. There's no $ shell passthrough for Claude Code.

Subagents

When Claude Code spawns a subagent, its work appears as a Subagent entry in the transcript with a View full transcript action, rather than being flattened into the parent turn.

Thread history

Claude Code threads in LinkCode read and write the same on-disk history the standalone claude CLI uses. A thread you start in LinkCode shows up in the claude CLI's own history, and a thread you started in the terminal can be opened and resumed from LinkCode. Stopping and resuming a thread is fully supported.

Long threads compact automatically. LinkCode shows the compaction in the transcript rather than silently dropping the earlier conversation.