Docs

Supported agents

A verified comparison of the coding agents LinkCode drives, covering setup, sign-in, approval behavior, models, reasoning effort, commands, and thread resume.

LinkCode drives five coding agents: Claude Code, Codex, OpenCode, Pi, and Grok Build. They are not interchangeable — each has its own setup path, sign-in model, and relationship to asking before it acts. This page is the single reference for those differences; every other page links back here instead of repeating the table.

Not every agent asks before acting

Claude Code, Codex, and Pi ask for permission and let you choose how strict that is. OpenCode asks only where your own OpenCode configuration says to — by default it runs unattended. Grok Build never asks and cannot be switched: its headless mode has no way to wait for an answer. If you're new to LinkCode, start with Claude Code.

Capability matrix

Getting the agentSign-inApproval policyModel pickerReasoning effortResume a stopped threadRich diffs for edits
Claude CodeDownload in-app, or your own installIn-app, or an account you add5 policiesCurated listLow → Max, plus UltracodeYesYes
CodexDownload in-app, or your own installIn-app, or an account you add3 policiesCurated listPer model, next turnYesYes
OpenCodeDownload in-app, or your own installOpenCode's own sign-in, or an account you addIts own agentsFrom your connected providersNoYesNo — plain text
PiDownload in-app; runs in-processIts own auth file/env vars, or an account you add3 policiesFrom Pi's provider registryPer model, Low → xHighYesNo — plain text
Grok BuildInstall the grok CLI yourselfThe grok CLI's own login, or an account you addNone — always unattendedGrok 4.5Low → HighNoNo — plain text

Verified against LinkCode v0.15.1.

Getting an agent onto your machine

LinkCode ships no agent binaries. The first time you pick an agent, it resolves one in this order:

  1. Managed — a copy LinkCode downloaded into its own store, pinned to the version this release was tested against. The first download always asks; later refreshes for an agent you already have happen quietly.
  2. Detected — an install already on your machine, found on the host's PATH or in the usual install locations and version-checked. If it isn't the paired version, the agent is marked Unverified: usable, but untested against this release.
  3. Neither — the agent shows Not installed, with a Download button. Grok Build has no managed download; install the grok CLI yourself.

Settings → Agents shows which of the three each agent is currently using.

Accounts and API keys

Agents authenticate through Settings → Providers, where you add the account once and connect it to whichever agents should use it: a Claude or ChatGPT subscription, a direct API key (Anthropic, OpenAI, xAI), a gateway (OpenRouter, Vercel AI Gateway, Cloudflare AI Gateway), or a custom endpoint. One account can back several agents; each agent uses one account at a time.

An agent with no account connected falls back to its own CLI login — the credentials you'd use if you ran it in a terminal. Claude Code and Codex can complete that login from inside LinkCode. Everything you enter is stored encrypted on the host and never written into an agent's own config files; LinkCode reads those files only to list and import existing history.

Choosing an agent

  • Claude Code is the most complete integration: five approval policies, a curated model list, live reasoning-effort switching, subagent transcripts, and history shared with the claude CLI. It's the safest default.
  • Codex is close behind: real approval prompts, per-model effort, $ shell passthrough, and it honors the sandbox in your own ~/.codex/config.toml until you pick a policy yourself.
  • OpenCode brings your own provider set and its own agent definitions, and it's the one whose permission behavior is driven by your OpenCode config rather than LinkCode.
  • Pi runs in-process across roughly two dozen providers, with LinkCode supplying the approval prompts Pi itself has no concept of.
  • Grok Build is the least integrated: no approvals, no resume, no commands.

For the reasoning behind approval policies vs. one-off permission prompts, see approvals. For how thread resume works in general, see threads.

On this page