GolemOpen source

Local-first · open source · Apache-2.0

Redact first.
Then choose where it goes.

Golem is a local layer between your tools and an LLM provider: a proxy and an MCP server. Request bodies are redacted before they are forwarded, nine upstream providers are supported, and savings figures say they are estimates.

  1. Redactbefore every other stage
  2. Answer locallyif it can
  3. Compresssituational
  4. Routeto a named upstream

Redaction before the model

The proxy redacts request bodies before compression and before anything is forwarded, at every compression level. It is best-effort pattern matching, not a guarantee, and the README lists what it does not cover.

Nine upstream providers

Anthropic, Azure Foundry, OpenRouter, OpenAI, Gemini, NVIDIA NIM, Ollama, llama.cpp and a custom endpoint. Anthropic is the default, and a local model is one more upstream rather than a special case.

A proxy and an MCP server

A proxy sits in front of model traffic, and an MCP server offers explicit tools: search, fetch, expand, a local coder, wiki read and write, and more. Several register only when what they need exists.

Local answers, quoted not generated

For an eligible question the proxy can answer from the project knowledge base without calling the model. The answer quotes retrieved text, carries a visible "verify independently" prefix, and can quote the wrong page.

Oversized output, kept not lost

When a tool result passes 12,000 characters, a hook swaps in a head and tail excerpt and a reference. The redacted original is stored locally and Claude can expand it on demand.

Observability that says "estimate"

Savings figures are estimates and are labelled that way. Compression is situational: on cached Anthropic traffic it buys close to nothing, and Golem says so rather than leading with a number.

What the portal adds

The portal supplies a team settings layer to the harness. The harness is complete without it, and nothing here is needed to run Golem.

One layer, in a fixed place

In the normal band the team layer ranks above a member's own settings and below the project, local and environment layers. A value marked important at the team resolves after those, and a member's own important declaration still beats it.

A team may only tighten

The policy table is default-deny. A team can force a risky boolean off but not on, force a protective boolean on but not off, and lower a numeric limit but not raise it. Anything that runs a command or sets a URL, endpoint, credential or path is refused, per key, by name.

As fresh as the last sync

Hooks read the team layer from a local cache, so a member's view is as current as their last golem team sync. If a team layer is skipped, policy is not in force for that run, and a member can still edit their own files.

Connectors to workplace tools

GitHub, Monday and Azure DevOps connect through the portal, with health tracking and admin alerts. Where a fact is not known, such as when an Azure DevOps token expires, the portal says so rather than implying health.

Three ways to run it

Nothing is held back from Free. Solo and Team exist because config that is hosted, and shared, has to live somewhere other than one disk.

Free

Local only

The whole harness on your machine, with config on your own disk. No account, no portal, nothing held back.

Free forever

Install Golem

Free, open source, and it runs entirely on your machine. No account needed.

curl -fsSL https://golem.run | sh

Installs to ~/.local/bin, or via npm if you have Node 22+.

Then

golem init
The script is served over HTTPS from the tagged release on GitHub, and that is the only thing vouching for it. There is no checksum or signature yet, so read it first if that matters to you. Full instructions.

Solo

Team of one

A team of one. Your own config hosted outside any single project or machine, and connections to your corporate tools.

$7 AUD / month

Team

3 or more

Three or more collaborators. Somewhere to host the config your team shares, beyond any one project, and connections to your corporate tools.

$5 AUD / seat / month, 3 seats minimum

The depth is in the repo: architecture, the spec, and what has actually been verified.

Read the README on GitHub