The Config Root IS the Identity.

CLAUDE_CONFIG_DIR=~/.claude-ac1 claude   # account 1: own auth/session/settings
    CLAUDE_CONFIG_DIR=~/.claude-ac2 claude   # account 2: own auth/session/settings
    claude                                    # default ~/.claude, untouched by either

Isolation Is Total by Default.

Each config dir starts blank. You opt back in to sharing by symlinking the stable, personal pieces — and leave the per-identity pieces real.

Symlink the Setup, Isolate the Identity.

# 1. create the new identity's root
    mkdir ~/.claude-ac1
    
    # 2. link the shared pieces back to the default setup
    for s in CLAUDE.md agents commands docs skills rules; do
      ln -s "$HOME/.claude/$s" "$HOME/.claude-ac1/$s"
    done
    # 3. credentials/settings/projects/plugins stay real (absent)
    # — created on first launch

Aliases Live Where Your Shell Lives.

A launcher alias per account means you never type the env var. But aliases only load from your interactive shell's rc file — and which file that is depends on your shell.

# one block per account, in the rc file for YOUR shell
    alias ac1claude='CLAUDE_CONFIG_DIR=~/.claude-ac1 claude'
    alias ac1claude-d='CLAUDE_CONFIG_DIR=~/.claude-ac1 claude --dangerously-skip-permissions'
    alias ac1claude-dr='CLAUDE_CONFIG_DIR=~/.claude-ac1 claude --dangerously-skip-permissions --resume'

Log In Once, Then Trust the Directory.

ac1claude      # launch with the ac1 config dir
    /login         # browser opens — sign in with that account

Login, API Key, or Cloud — One Identity, Many Auth Paths.

The credential lives inside the config dir, so each account authenticates differently:

// ~/.claude-ac1/settings.json        ~/.claude-ac2/settings.json
    { "env": {                            { "env": {
      "ANTHROPIC_API_KEY": "sk-ant-…"        "CLAUDE_CODE_USE_BEDROCK": "1"
    } }                                    } }

Point the Harness at GLM.

Claude Code speaks the Anthropic API — any provider that speaks it back drops in. These env vars redirect the harness, tiers remapped:

The X_ Prefix Toggle.

Keep every provider's vars in global settings.json; prefix the inactive set with X_ so Claude Code ignores them. Activate one per project in .claude/settings.local.json (gitignored, overrides global) — GLM here, Anthropic login next repo, Bedrock in a third.

// prefix = INERT (global) · drop X_ = LIVE (project-local, gitignored)
    {
      "env": {
        "X_ANTHROPIC_BASE_URL":   "https://api.z.ai/api/anthropic",
        "X_ANTHROPIC_AUTH_TOKEN": "…"
      }
    }

OpenCode Alongside Claude Code.

Claude Code is one vendor — predictable, polished, capped on heavy days. OpenCode is the open-source counterpart: one UI over 75+ providers, switchable mid-session.

OPENCODE_CONFIG_DIR — OpenCode's Multi-Account Lever.

Claude has CLAUDE_CONFIG_DIR; OpenCode matches it — one env var for the file, one for the dir:

alias oc1='OPENCODE_CONFIG=~/.config/opencode-ac1/opencode.jsonc \
               OPENCODE_CONFIG_DIR=~/.config/opencode-ac1 opencode'
    alias oc2='OPENCODE_CONFIG=~/.config/opencode-ac2/opencode.jsonc \
               OPENCODE_CONFIG_DIR=~/.config/opencode-ac2 opencode'

Keep Zen and GLM — Switch by Flipping a Default.

provider is a map, not a single choice — predefine Zen and GLM in one file, switch by changing model (or /connect / /models). No re-auth each time.

// opencode.json — Zen (built-in) + your GLM, one active
    {
      "model": "zai/glm-4.6",                    // ← flip to "opencode/gpt-5.5" for Zen
      "provider": {
        "zai": {
          "npm": "@ai-sdk/openai-compatible",    // protocol adapter — built-ins (Zen) need only a key
          "options": { "baseURL": "https://open.bigmodel.cn/api/paas/v4/", "apiKey": "{env:ZAI_API_KEY}" },
          "models": { "glm-4.6": { "name": "GLM-4.6" } }
        }
        // "opencode" (Zen) is built-in — added by /connect, billed per request
      }
    }

Share Context, Don't Duplicate It.

You already wrote your standards once — in CLAUDE.md and rules/. OpenCode should read the same files, not a hand-maintained copy. Its config takes an instructions array of paths and globs, loaded at every session start:

// ~/.config/opencode/opencode.jsonc
    {
      "$schema": "https://opencode.ai/config.json",
      "instructions": [
        "~/.claude/CLAUDE.md",
        "$HOME/.claude/rules/*.md"
      ]
    }

Skills Auto-Port. Agents Don't.

OpenCode reads Claude-compatible locations on purpose, so most of your setup crosses over with zero config:

Which Repo Rules Auto-Load — and Which Don't.

Run inside a project and OpenCode auto-discovers a fixed set of filenames, traversing up from the working directory. Everything else is invisible without a glob.

A Directory, an Env Var, a Credential — At Every Scale.

The same harness shape repeats. Change the directory, change the identity — whether that identity is an account or a provider:

Tune For Your Stack.