Magpie vs CC Switch: Why I Switched to This AI Model Gateway for Multi-Provider Routing

After using CC Switch for a while, the author discovered it silently dropped parts of ~/.claude/settings.json when switching providers: status-bar config, 13 hooks, enabledPlugins, extraKnownMarketplaces, and specific permissions.deny rules vanished, while an unknown environment variable appeared. The root cause was CC Switch's snapshot-based provider storage: adding a provider copied the entire current config, and switching wrote that snapshot back. From version 3.16.5, CC Switch attempted to protect shared sections (hooks, enabledPlugins, env) by re-reading them from the live settings.json before each switch, but this protection only applied to providers with "common config" enabled. The author's sole provider without that flag was DeepSeek, so its stale snapshot overwrote the live file.

To prevent further loss, the author built a skill, two hooks, and a reconciliation script guarding settings.json. Around the same time, they discovered magpie (v0.1.1564) and decided to trial it. This article records that experience.

No More Manual Provider Switching

Installing magpie adds a local gateway at 127.0.0.1:3425 that translates between OpenAI Chat Completions/Responses, Anthropic Messages, and Google Gemini API formats. CC Switch also does protocol translation, but its supported directions depend on the connected clients and upstreams.

Magpie aggregates all providers' models into a single list visible inside Claude Code's model picker. For each of Claude Code's four tiers (opus, sonnet, haiku, fable) magpie provides a dropdown; each tier can point to the same model or different ones (e.g., expensive model for opus, fast model for haiku), and child agents can get their own assignments. Adding providers is simple: built-in presets cover common services, and custom providers can be defined. Providers requiring extra login or request logic (browser auth, device-code flow, periodic token refresh) are handled via plugins loadable from npm, git repos, or local folders.

Automatic Provider and Account Selection

Enabling "auto-create route groups" merges same-named models from multiple providers into a group (e.g., group/auto-deepseek-v4-1-flash containing OpenCode Go and Volcano Engine). The current docs denote auto groups as auto/<model-id>; the older group/auto-… naming remains compatible. Manual groups can combine different models. When an agent selects a group, magpie routes requests according to the group's strategy.

Routing strategies (per provider account/key) include:

Smart (default) : among accounts with remaining quota, pick the one whose reset window ends soonest.

Sequential : follow the configured order; move to the next only when the current fails.

Round-robin : rotate to the next account/key on each re-selection.

Least-used : prefer accounts with lowest usage ratio (quota-aware) or recent token count (plain keys).

Use-up-before-reset : prioritize accounts that can consume the most per hour before their window resets.

On the default smart strategy, if two accounts have ample quota but one resets tomorrow and the other next week, the former is used first. Accounts with only a 5-hour window (e.g., Claude Enterprise) are treated as resetting in 5 hours and placed before weekly-reset accounts. Accounts at ≥90% quota usage are deprioritized.

Additional "when rate-limited" setting: "stay" or "move to end". Exhausted accounts pause until reset; billing-suspended accounts wait 30 minutes; vendor-rate-limited accounts honor Retry-After; erroring accounts back off with increasing delays. Failover only occurs before the first response chunk reaches the agent — mid-stream switching is not supported. CC Switch has similar failover; magpie adds proactive scheduling of account order.

Session persistence controls how long a selection sticks. A "turn" starts with the user message and ends when the agent finishes replying. Options:

Auto (default) : hold within a turn; across turns, evaluate whether the vendor's cached context is worth keeping.

Entire session : reuse the same account as long as it responds normally.

Within one turn : stable for the current turn's tool calls; re-select on next user message.

Off : re-select on every request.

Exceptions: if the current account cannot answer, or context exceeds the model's capacity, magpie switches immediately.

Request-level routing rules can match on token length, presence of images, reasoning requirement, originating agent, context-compression phase, time windows, weekdays, and folder inclusion/exclusion. Examples: route image inputs to a vision-capable model; delegate web search to a designated search API/model; detect and terminate agent loops (repetitive reasoning/output) by returning a retryable error.

Intent routing uses a classifier model (configurable, default 8s timeout, max 60s) that sees the latest user message plus truncated previous user/assistant pair to understand follow-ups like "continue" or "fix this". System prompts, thinking traces, tool calls/results, and Claude Code's system-reminder are stripped. On classifier error/timeout/invalid output, the previous turn's intent is reused if still a candidate; otherwise fall back to other rules. Classifier usage appears on the usage page; local models via Ollama/LM Studio can serve as classifiers.

Route groups need not be built from scratch. When no groups exist, three templates are offered: "Hard tasks to strong models, easy tasks to cheap models", "Cheap first" (escalate to strong only when cheap fails or high reasoning needed), and "Never stuck" (each provider's strongest model backs up the others). Groups can also be created/adjusted via CLI.

Observability: See What It Does

Magpie surfaces its modifications in each agent's model config UI (e.g., shows which fields in ~/.claude/settings.json were changed). The usage page breaks down tokens, cache hit rates, and estimated cost per model and account, with per-account detail showing attempts, responses, failures, used quota, and reset time. The route page's live stream lists every request proxied through magpie: which account responded, why it was chosen, and each attempt. Drilling down shows real-time context size, cache hits, latency, and cost. The gateway page displays structured raw messages of recent calls.

Fine-Grained Parameters & Safe Config Management

Provider parameters (context window, max output, compression threshold) can be overridden per model (e.g., deepseek-v4-flash=272k). Empty fields fall back to vendor/models.dev data (context/max-output) or the global "long conversation" setting (compression threshold). Each account can have its own network connection settings (inheriting or overriding provider-level config) and concurrency limit (unlimited by default).

Connecting an agent requires editing its config — a risky operation. Magpie reads the original file first, updates only the fields needed for the gateway, preserves comments, key order, indentation, and writes atomically. For Claude Code it touches the env section: ANTHROPIC_BASE_URL, ANTHROPIC_AUTH_TOKEN, and the four tier model variables. On the author's machine ANTHROPIC_BASE_URL points to 127.0.0.1:3425 and the token is the string magpie. For Codex it adds a [model_providers.magpie] table in ~/.codex/config.toml and a separate magpie-models.json; the official docs state it does not disturb the existing ChatGPT login.

Magpie also creates new config directories when needed (e.g., a Claude-3p directory for Claude Desktop third-party providers). It records every change: applied.json stores the last applied fields and timestamp per agent; stash.json keeps pre-connection values and state for rollback. Reverting is one click ("Disconnect and Restore") or via magpie <agent> default, which removes magpie's fields and restores originals including any pre-existing ANTHROPIC_BASE_URL.

Noteworthy Details

UI : The author finds magpie's interface more comfortable for extended use than CC Switch's.

Tray : Single icon shows remaining quota on hover; click opens quick panel or main window. For subscriptions with reset cards, the card displays remaining cards and expiry; "auto-use reset cards" automatically attempts quota reset when a card is available.

Window preheating for 5-hour-window subscriptions: "Preheat on window reset" sends a tiny request right after reset to start the next window early; "Daily scheduled preheat" launches windows at a set time each day so morning sessions may already have a running window. If work spans a reset point, both windows' quotas become usable sequentially (each window's cap still set by vendor).

Privacy : Dedicated page with masking/redaction options.

Conclusion

CC Switch 4.0 fixed the config-writeback issue and added model aggregation. However, for the author's multi-provider, multi-account workflow, magpie's automatic routing and quota-aware scheduling fit better, so they have no plans to switch back.

Magpie evolves quickly; adapting to new settings takes time. If CC Switch already meets your needs, there's no reason to migrate just because of this article.

Magpie also offers a CLI/TUI version ( magpie tui) for terminal-only users; see the official Chinese docs for details. CC Switch updates can be tracked at its official repository.