Connecting Claude Code to Parel
You use Claude Code daily. The Anthropic subscription is $200/month, but 80% of
your day is simple file navigation and edits. Solution: keep Claude Code as it
is, route the model brain through Parel. Use tm_qwen3coder
($0.18/M tokens) for routine work, claude-opus-4-7 only for hard
refactors. Setup takes 30 seconds.
ANTHROPIC_BASE_URL + ANTHROPIC_AUTH_TOKEN) and
Claude Code's /model picker shows Parel's 100+ models. Command,
IDE integration, MCP setup all stay the same.
1. Install the Parel CLI (30 seconds)
# macOS / Linux
curl -fsSL https://parel.cloud/install.sh | sh
# Windows (PowerShell)
iwr -useb https://parel.cloud/install.ps1 | iex
# Brew alternative
brew install parel-cloud/tap/parel
# 30 seconds, doesn't touch the claude command 2. Login + Claude Code init
parel auth login # opens a browser for OAuth
parel claude-code init # writes env vars to ~/.zshrc or $PROFILE
parel claude-code status # verify: ANTHROPIC_BASE_URL set?
# Open a fresh terminal and continue parel claude-code init writes 4 env lines to your shell profile
(idempotent, running it twice doesn't duplicate). To roll back run
parel claude-code uninstall.
3. Smoke test
# Verify with curl before opening Claude Code
curl https://api.parel.cloud/anthropic/v1/messages \
-H "x-api-key: $ANTHROPIC_AUTH_TOKEN" \
-H "anthropic-version: 2023-06-01" \
-H "content-type: application/json" \
-d '{
"model": "claude-opus-4-7",
"max_tokens": 64,
"messages": [{"role": "user", "content": "One sentence: which model are you?"}]
}'
# Expected: 200 + Claude's reply
# 401 = ANTHROPIC_API_KEY must be empty (empty string)
# 404 = wrong model name (claude-opus-4-7, not "4.7") 4. Three lanes in the model picker
Open Claude Code (claude), type /model. Parel's three
lanes appear:
| Lane | Example | $/M tokens | For |
|---|---|---|---|
| Showcase | claude-opus-4-7, gpt-5.4 | $15-$30 | Hard refactor, architectural decisions |
| Showcase | claude-haiku-4-5, gemini-3-flash | $1-$3 | Quick edits, file navigation |
| Instant | tm_qwen3coder, tm_qwen3max | $0.18-$0.42 | Routine coding, most work |
| Instant | tm_llama3370b, tm_gpt4omini | $0.18-$0.30 | Test writing, simple edits |
| BYOM | byom-d2k7x9 | $0.22-$3.20/hr | Your fine-tune |
# Claude Code's /model picker now shows 3 lanes:
#
# claude-opus-4-7 (showcase — hard refactors, complex debugging)
# tm_qwen3coder (instant — routine coding, file navigation)
# tm_gpt4omini (instant — fast hot-fixes, simple edits)
# byom-DEPLOY_ID (BYOM — your fine-tune)
#
# Example flow:
# 1. claude --model tm_qwen3coder → plan a 50-file refactor
# 2. /model claude-opus-4-7 → review the critical diff first
# 3. /model tm_qwen3coder → apply to the other 49 files 5. Manual install (without the CLI)
Skip the CLI by writing the four env vars yourself. ANTHROPIC_API_KEY
must be an empty string; otherwise the Anthropic login picker isn't bypassed
and Claude Code goes to Anthropic instead of Parel.
# If you don't want to install the CLI, write env vars yourself
# ~/.zshrc, ~/.bash_profile or $PROFILE
export ANTHROPIC_AUTH_TOKEN="parel_pk_xxxxxxxxxxxx" # Parel key, app.parel.cloud/api-keys
export ANTHROPIC_API_KEY="" # MUST be empty
export ANTHROPIC_BASE_URL="https://api.parel.cloud/anthropic"
export ANTHROPIC_CUSTOM_MODEL_OPTION="parel/claude-opus-4-7" Tool-use compatibility
Claude Code's productivity comes largely from tool-use (file read/write, shell, MCP). Tool-use across lanes:
| Model | Tool-use | Note |
|---|---|---|
| claude-* | Full | Native, Parel passes through |
| tm_qwen3coder | Full | Function-call → Anthropic tool_use mapping |
| tm_gpt4omini | Full | OpenAI function → Anthropic tool_use |
| BYOM (vLLM) | Full | Requires --enable-auto-tool-choice |
| BYOM (TGI) | Limited | No native function-call; "JSON only" prompting recommended |
Decision
Ship
Smoke 200, picker shows Parel entries, agentic flow (file + shell + commit)
works on tm_qwen3coder. Cancel the subscription, keep going on
a Parel key.
Hybrid
Routine on tm_qwen3coder, hard refactors on
claude-opus-4-7. The $200/month subscription becomes a variable
Parel bill — most teams see a 60-80% reduction.
Stop
If you only have one senior workflow doing hard refactors and debugging, the Anthropic showcase is enough. Until you need BYOM or open models, stay on the direct subscription.