Goose Lets You Swap Models. It Will Not Swap Out the Bill

Herdr last night was a multiplexer for agents you already pay for. Goose is the other itch. You want Claude on the hard problem, a cheap model on the grep, maybe a local one when the laptop is on a train, and you do not want four desktop apps.

Gigazine’s 12 September writeup is the one that circulated. Docs at goose-docs.ai. Code at github.com/aaif-goose/goose. The pitch is an open-source agent that is not glued to one vendor: OpenAI, Claude, Gemini, local LLMs, desktop, CLI, API, MCP tools. It edits files. It runs commands. It is closer to “a process that is allowed to touch the repo” than to a chat window that pastes snippets.

I am going to tell you what the writeup actually shows, what it rhymes with in Copilot HydraFusion, and why Spotify spent the same week proving that the expensive part is still context.

What you install

Goose is a Block project in the open. The GitHub line Gigazine quotes: an extensible agent that goes beyond code suggestions, install, execute, edit, test, with any LLM. Desktop app for people who want a window. CLI for people who already live in a terminal. API if you are wiring it into something else.

The desktop path in that article starts with file edits. You point it at a file, it changes the file. That sounds obvious until you remember how many “agents” still copy a diff into Slack and wait for you. If Goose is doing the write, your git status is the audit log. Keep it that way. Do not give it a clean tree and a bored afternoon.

MCP is the integration story. Tools show up as servers. Browser, docs, internal APIs, whatever you are willing to expose. That is the same shape as every serious agent in 2026. It is also the same attack surface we keep writing about when a sandbox lets the model turn itself off. Stick to tools you would run yourself.

Provider switching is the feature people will blog. Claude for the refactor, Gemini for the long PDF, a local model for the commit message. The config is the product. If switching models takes a new install, Goose failed. If it is a dropdown or a YAML key, Goose did the job Herdr does not try to do. Herdr keeps the session alive. Goose is the session.

Build from the repo if you do not like vendor installers. Read the docs URL, not a random gist. I am not pasting a curl | sh from a secondary blog for the same reason I would not do it for Herdr.

A day in it, without a sales demo

Pick one repo you can throw away. Open the CLI. Set a provider you already have a key for. Ask it to list the test command and run it. If it cannot run tests, it is a chatbot with extra steps.

Then ask it to edit one function and show you the diff before it writes. If the desktop app writes silently, turn on whatever dry-run exists or stop using the desktop app. The CLI plus git is enough.

Third, add one MCP tool you understand. A filesystem server on that repo is enough. Not your password manager. Not prod.

Fourth, switch provider on the same thread if the product lets you. If the thread dies, the switching story is marketing. If the thread lives and the new model is lost, that is expected. Models do not share a brain. Goose is a harness.

Pricing: Goose the binary is free. The tokens are not. Local models are free until you count the GPU. This is not a way to avoid Anthropic. It is a way to route around Anthropic for the work Anthropic should not see.

Spotify already told you where the money goes

The same Gigazine day had Spotify engineers cutting Claude Code token use by about 90%. Different tool. Same physics. Mazmanov, in that piece, used AiKA Modes on Portal by Spotify: declarative instructions, model choice, temperature, MCP tools, called as one agent from a CLI or API. The environment is temporary. No always-on agent server.

The trick with a name is bulk-reader. Claude Code should not swallow a thousand-line file because it is curious. It hands the file and the question to bulk-reader. bulk-reader reads, returns a short answer, Claude stays on a diet. High-end reasoning is for the part that needs it. The rest is a cheaper pass.

If you install Goose and then dump the monorepo into context, you have rebuilt the bill Spotify just killed. Use a summarizer tool. Use a smaller model to grep. Save Claude for the function you cannot read yourself.

Ninety percent is their number, on their platform, in a Japanese writeup of an English engineering story. The architecture is not Goose-specific. Any harness that cannot intercept “read this whole file” will eventually need a bulk-reader.

Where this sits on the shelf

Herdr: keep Claude Code and Codex alive when you close the lid. Goose: be the agent, talk to whoever’s API, touch files. HydraFusion: Microsoft’s cost knob on Copilot. Spotify Portal: an internal version of the same idea with receipts.

Goose wins if you hate being locked to one chat product and you are willing to own MCP config. It loses if you wanted a managed pair programmer that already knows your org. Block is not going to staff your on-call.

Abacus and the all-in-one web workspaces are a different aisle. They sell breadth and a monthly price. Goose sells a process on your machine. Do not mix the shopping lists.

I would run Goose on a throwaway clone this week, with one key, with git dirty on purpose so you see the writes. I would not point it at the company monorepo on day one, and I would not call it a Herdr replacement. Different job. Same rule: the model is the expensive intern. The harness is how you stop the intern from reading the entire filing cabinet.

What I would configure, and what I would refuse

One provider key in a file that is not in the repo. If Goose wants a .env at the project root, gitignore it on the first commit, not the second. Local models via whatever runtime you already use. Do not install a second GPU stack “for the agent.”

Permissions: write in the clone, network only to the model API and the one MCP you added. If the desktop app asks for blanket disk, use the CLI. If the CLI wants sudo, stop.

A small bulk-reader of your own is a script that takes a path and a question and prints two hundred words. Wire it as MCP. That is the Spotify lesson without Portal. Goose does not owe you AiKA Modes. You owe yourself a tool that says no to whole-file dumps.

Refuse: production kubeconfig, CI tokens, anything that can open a PR unattended on day one. Agents that can git push are how you get a Friday incident with a polite commit message. Open a PR by hand until you have watched a week of diffs.

Refuse also the fantasy that swapping models is a quality knob. Claude to Gemini is a different intern, not a smarter one. Goose makes the swap cheap. It does not make the intern cheap. Spotify’s 90% was not a model swap. It was refusing to feed the intern the warehouse.

If Goose’s docs and the GitHub repo disagree, believe the repo for flags and the docs for intent. Secondary blogs will date out. Gigazine is how this got on the calendar. It is not how you debug a provider plugin at 1 a.m.

That is the review. Free harness, paid tokens, useful if you will starve the context, noisy if you will not. Install it like you would install a new grep, not like you would hire a colleague.

A last practical note. Herdr and Goose can sit on the same machine without being friends. Herdr owns the pane. Goose owns the process in the pane. If you run Goose inside Herdr, you get persistence for a harness that already writes files. That is convenient and slightly more dangerous, because a detached session will keep editing after you walk away. Set a timeout. Watch the first hour. If it loops on a test it cannot pass, kill it. An open-source agent with no adult in the room is just a loop with an API key.

I am not ranking Goose against Claude Code. Claude Code is a product with a default model. Goose is a sleeve. People who want a sleeve should try this week’s build, read the GitHub issues before the homepage, and keep Spotify’s bulk-reader in their head even if they never work at Spotify. People who want a pair programmer should pay Anthropic and stop shopping for a personality. Both groups should count tokens. The 90% story exists because somebody did.