magpie not working? 401 and 405 errors and fixes (Nov 2026)

A running session keeps the model it started with, as the official FAQ says. Start a new session, and for Codex restart the Codex app too. Agents read their config when they start, so the change applies to the next session. To put an agent back the way it was, pick one of its own models in the picker, or run magpie followed by the agent name and default, for example magpie claude default.

This is the 401 text quoted in magpie's Docker docs: the key an agent sent from outside the container is not an enabled gateway key (with LAN sharing, requests from other machines need a gateway key as well). The docs name one case: after upgrading from an earlier image, an agent outside the container still sends magpie or another key that is not a gateway key. The fix: make a key under Gateway → Gateway keys, or with magpie gateway-key add and a name (in Docker, put docker exec magpie in front), which prints a new sk-magpie-key-…, and put it in the agent as its API key. On local loopback any token works.

The title of issue #425 begins ZCode provider returns 405 Method Not Allowed, and the issue is closed (not planned). The author's reply on 2026-10-01 gave the principle that magpie does not rewrite your prompts and does not pose as the official client. Since v0.1.590 this error message says ZCode only accepts requests from its own app and suggests adding another provider to the group. If you follow it and add a provider that works with an API key, QCode offers GLM models such as glm-5.3 on both the Anthropic and the OpenAI Chat interface; see /models for each model.

No, as long as the anthropic line has a check mark. QCode serves Claude models only on the Anthropic Messages endpoint; when the first model is a Claude model, the test sends it to chat and responses as well and gets a 400 by design, reading Model 'claude-opus-5-5' is not available on this endpoint. To get a check mark on all three lines, put a GPT model such as gpt-6.1-terra first when you list models with magpie provider models, as QCode tested on 2026-11-10; you can also check GPT by running magpie codex qcode/gpt-6.1-sol and sending a real request.

In QCode's test, without models= when adding the provider, magpie exposes only the first 24 models. Writing the provider id and model IDs after magpie provider models replaces the whole list with yours; to expose just one more, the reference uses +model-id (and -model-id takes one out), leaving the rest as they are. For a messy list, the author's answer on X on 2026-11-05 is to create a routing group and pick only that group in the agent; the reference also describes an Only through routing groups switch on a provider that keeps the whole provider for routing groups.

The cause in the official FAQ: a proxy app (Surge, Clash, mihomo and the like, with system proxy, TUN or enhanced mode on) has taken the agent's request to http://127.0.0.1:3425 and sent it somewhere else, so it never reaches magpie. The official fix: add two rules that send local addresses direct, ahead of your other rules, IP-CIDR,127.0.0.0/8,DIRECT,no-resolve and IP-CIDR6,::1/128,DIRECT,no-resolve, then have the agent send a new request.