magpie Gained 3,500 Stars in a Week While Its Code Grew 21x Past Its Own README

yetone/magpie was created on September 23 and had 3,526 stars when we checked tonight, seven days later. Our own trending snapshots caught it at 2,999 stars at 04:00 UTC today and 3,522 by 13:00, so it added 523 in nine hours, which is acceleration rather than a launch spike fading out.

The pitch is one screen that shows which model every AI coding agent on your machine is set to, plus a local gateway so Claude Code can run on Kimi or Codex on DeepSeek. We read the source, built it, pointed it at a fake vendor we wrote for the purpose, and let it edit a set of throwaway config files. No real keys or real configs were involved at any point.

What it actually is

Two things in one Go binary. The first is a config editor that knows where 29 agents keep their settings, from ~/.claude/settings.json to ~/.codex/config.toml to OpenCode's opencode.jsonc. The second is a gateway on 127.0.0.1:3425 that speaks OpenAI chat completions, OpenAI Responses and Anthropic Messages, and translates between them so any agent can use any vendor.

magpie provider add deepseek sk-...      # a preset needs only the key
magpie claude deepseek/deepseek-chat     # rewire Claude Code
magpie codex effort high
magpie save work                          # snapshot every agent
magpie serve                              # the gateway on its own

The README says the config editor plus gateway "is the whole app". The help text in the build we tested also lists encrypted backups, sync over WebDAV and S3, switching between multiple Claude and ChatGPT subscriptions, a daily check-in for WorkBuddy accounts, usage and cost reports, and an image generation MCP server. That is a lot of surface for a week-old project, and it matters for the next section.

The README froze on day one

The README promises "one small binary", 7 MB for the terminal build and under 15 MB with the desktop app. We built commit 3bec6c0 with the repo's own Makefile targets on an M4 Pro, then built the very first commit from September 22 the same way.

Build	Claimed	We measured
First commit, terminal build (5,089 lines of Go)	7 MB	7.4 MB
Today, terminal build (make cli)	7 MB	22.2 MB
Today, desktop build (make build)	under 15 MB	26.1 MB
Today, desktop build, gzip -9	under 15 MB	10.2 MB

The claim was true when it was written. Since then the non-test Go has grown from 5,089 lines to 108,408, a 21x increase in eight days, and the binary tripled. Nobody is being dishonest here. The number simply stopped being updated, and that tells you something about the pace: 847 commits since September 22, with 147 on September 29 and 149 more today.

We let it edit real-shaped config files

We set HOME to a scratch directory and wrote three configs the way a person does: a Codex TOML with inline comments, an OpenCode JSONC with // comments and a trailing comma, and a Claude Code settings file with four-space indents. Then we added our fake vendor and pointed all three agents at it.

The strongest claim is that edits are surgical, "only the one key you change is touched". Formatting really did survive: every comment on a line magpie did not rewrite was still there, the trailing comma stayed, the indentation stayed. The scope did not match the claim.

One command: set the model	What changed
magpie claude mock/mock-1	model plus 9 new env entries, including ANTHROPIC_BASE_URL and all five model aliases
magpie codex mock/mock-1	model, 3 new top-level keys and a new 4-key [model_providers.magpie] table; the inline # the one I trust comment on the model line was dropped
magpie opencode mock/mock-1	model plus a 16-line provider block

That wiring is how the product works: Claude Code cannot use DeepSeek without its base URL pointing somewhere else. But "one key" undersells it. Undoing it is where we would push back hardest. magpie claude default does not restore the opus we had. It deletes the model key and leaves an empty "env": {} behind. Saving a profile first with magpie save original and restoring with magpie use original did bring every value back, but the Codex comment stayed gone and 23 lines of magpie wiring remained across the three files.

The gateway is the good part

We wrote a 36-line Python server that impersonates an OpenAI-compatible vendor and logs every request, then sent Anthropic-format requests to magpie and compared both sides.

• A streaming tool call came back as correct Anthropic events: tool_use block start, two input_json_delta chunks, stop_reason: tool_use, with token usage carried through.
• A multi-turn conversation with a tool_result arrived at the vendor as a proper OpenAI role: tool message with the matching tool_call_id, and the system prompt became a system message.
• Across 200 requests each, the gateway added about 0.7 ms: a median of 0.98 ms through its Anthropic endpoint against 0.30 ms calling the mock directly. Against a real model's response time, that is nothing.
• All 30 Go test packages passed, in 153 seconds including compilation.

What it does not do

• It does not restore your previous setting when you go back to "default". Save a profile before your first change.
• It does not authenticate local callers of the gateway.
• It does not keep inline comments on lines it rewrites.
• It has no GitHub releases yet. You build from source, or install from usemagpie.ai, which we did not test.
• We did not test the desktop app, subscription sharing or any real vendor, because each would have meant signing in with real accounts. Those remain the maintainer's claims.

Who made it, and does that matter

yetone wrote 770 of the commits. The next-busiest contributor has 15, and no one else has more than 5. yetone has shipped before: cumora, from August, has 3,917 stars. But a 108,000-line Go codebase that rewrites the config of every coding agent you own is far too big to vendor if one person stops. The dependency list is modest for its size, with 14 direct modules including Wails for the desktop app and a pure-Go SQLite.

Verdict

If you already juggle three agents across two vendors and hand-edit ANTHROPIC_BASE_URL, the gateway alone is worth running. It translated everything we threw at it correctly and costs under a millisecond. Run magpie save before anything else and put your configs in git. If you use one agent with one vendor, wait. You would be taking on a 22 MB binary that changes 150 times a day to avoid editing one line. What would change our minds is a release cadence slow enough that the README can keep up.

Your turn

How many coding agents do you actually have configured on your machine right now, and how many different vendors are behind them? If the answer is one and one, tell us what would make a tool like this worth it. For another agent tool whose stars ran ahead of its own install instructions, read our look at fast-jev-compaction, where the README's npm line returned a 404.