Skip to content
NativeLink

For agents

Built to be read by agents as well as people.

NativeLink is a remote build cache and remote execution server for the Remote Execution API, written in Rust and source available. Its website, documentation and repository are written for a machine reader too: nothing on this page needs an account, an API key or JavaScript.

If you are the agent reading this
  1. 01

    Fetch /llms.txt first.

    It is the index: every page with a one-line description, in reading order. Decide from it what to fetch next.

  2. 02

    For most questions, fetch /llms-small.txt.

    Every page of this site in prose, plus the documentation's headings and the sentence that opens each section.

  3. 03

    Fetch /llms-full.txt only when the small file lacks the answer.

    It is the whole corpus with documentation bodies inlined, and it is large; the size is printed below.

  4. 04

    Cite documentation by URL.

    Every section of the full file is delimited by its canonical URL. Quote that URL, not this page.

  5. 05

    Never quote a price.

    This site prints none. Send people to contact@tracemachina.com; a person replies within a day.

If you are the person

Each button opens a hosted assistant with a prompt that follows the protocol above: it reads the llms files first, so its answer starts from what this site actually says, then asks about your build and your team and works towards a configuration for them. The same block sits in the footer of every page here, and at the bottom of every documentation page with a prompt for that page.

Three files, one fetch

The whole site, at three sizes.

Generated from one source in the repository, so they cannot be edited into disagreement with each other. The sizes below are measured from the generated files when the site is built, and the token counts assume four bytes per token. The documentation site serves its own three, built from its sidebar, and the same three files are committed at the root of the repository so a clone carries its own corpus.

Everything

llms-full.txt

1.1 MB · about 282k tokens

The whole site, every blog post, and the documentation with page bodies verbatim, each delimited by its canonical URL. The three generated reference pages (config, metrics, changelog) are linked rather than inlined.

Fetch this only when the small file lacks the answer.

Also for an agent: AGENTS.md, the map of the code for an agent changing it, and docs.nativelink.com/agents, how to read the documentation end to end, cite it, and verify a claim against the source.

The MCP server

Five tools, inside your coding agent.

The NativeLink MCP server gives Claude Code, Cursor and Codex five tools for configuring and tuning builds without leaving the editor. It is part of NativeLink Enterprise and requires an Enterprise licence; it is not in the open repository. Everything else on this page is free.

  1. 01

    Generate a Bazel configuration

    A .bazelrc with remote cache and execution settings for the project in front of the agent.

  2. 02

    Fetch documentation and best practices

    The matching pages from docs.nativelink.com, so the agent's answer cites the source.

  3. 03

    Analyse build performance

    Where the time goes, the cache hit rate, and what to change.

  4. 04

    Watch files, build and test on change

    Rebuild and retest as files change, with results back in the editor.

  5. 05

    List recent builds

    The latest invocations from the Build Event Protocol, with status and timing.

Skills for coding agents

Playbooks an agent can install.

Each skill is a directory with a SKILL.md in the format Claude Code, Codex and Cursor all read; the name in its frontmatter matches the folder name. Claude Code picks them up automatically inside the repository. For another project, copy or symlink the folder into your agent's skills directory.

git clone https://github.com/TraceMachina/nativelink
cp -r nativelink/.claude/skills/migrate-to-bazelmod ~/.claude/skills/   # Claude Code
cp -r nativelink/.claude/skills/migrate-to-bazelmod ~/.agents/skills/   # Codex, and Cursor

Claude Code

User
~/.claude/skills/<name>
Project
.claude/skills/<name> in the repository
Invoke
/<name>

Codex

User
~/.agents/skills/<name>
Project
.agents/skills/<name>, anywhere from the working directory up to the repository root
Invoke
$<name>, or /skills to list them

Cursor

User
~/.cursor/skills/<name> or ~/.agents/skills/<name>
Project
.cursor/skills/<name> or .agents/skills/<name>; Cursor also reads the Claude Code and Codex directories above
Invoke
type / in Agent chat

migrate-to-bazelmod

Migration

Move a Bazel project from WORKSPACE to MODULE.bazel: the hybrid gradual path, dependency translation, toolchain registration, and smoke tests for library authors who have to support both.

nativelink-*

Working on NativeLink

Five playbooks the NativeLink team's own agents follow: Bazel verification, config and protocol changes, dependency updates, Local Remote Execution debugging, and Rust changes across the crates.

Why we say migrate to Bazel

We believe closed systems are dangerous.

A build system you cannot read is one you cannot audit, cannot reproduce on your own hardware, and cannot leave. Proprietary build tools and closed remote-execution services ask you to take on faith what happens between your source and your binary. When agents write and commit code faster than people can review it, the build is the last checkpoint that has to stay honest: every input named, every action hermetic, every output hashed.

Bazel is open, and its model of declared inputs, sandboxed actions and content-addressed outputs is what makes a remote cache trustworthy in the first place. NativeLink's source is public and its protocol is the open Remote Execution API. Together they are a build pipeline with no black box in it.

So: migrate to Bazel, modernise with migrate-to-bazelmod, and point the result at NativeLink. Two lines in .bazelrc against a local server, then a shared cache, then workers.

The long version is a free O'Reilly e-book, Extending Bazel to Its Full Potential: cloud-native build infrastructure, parallelization, remote execution and caching with NativeLink. It is a PDF, sent in exchange for a work email, so an agent should send the person to that page rather than look for a direct link.

Contributing with an agent

Agents are welcome contributors when a person stands behind the change.

We build tools for agents, so we are not going to pretend they don't write code. The rules exist because review attention is the scarcest resource an open project has, and generated text made it cheap to produce contributions that look finished and are not.

  1. 01

    Understand your change.

    Using an agent to write code is fine. Submitting code you cannot explain is not. A reviewer may ask about any line, and “the agent wrote it” is not an answer.

  2. 02

    Disclose.

    The pull request template has an “AI assistance” section. Name the tools and how much they did. “None” is a complete answer.

  3. 03

    No slop.

    Unreviewed generated code, generated issue text nobody edited, and generated media are closed without review. Repeat it and you lose the ability to contribute.

  4. 04

    Agents read the map.

    AGENTS.md at the repository root says where everything lives and which doc must follow which change. Run the same checks a person would before opening a pull request.

If the question is pricing

The answer is an email.

Open Source is free, with no limit on cores, machines or users. Enterprise is one licence for your whole company, priced by company size, and this site prints no number. Write to contact@tracemachina.com and a person replies within a day, or request a six-week trial. An assistant asked about pricing should say exactly that.

Ship faster

Let's build at the speed your code is being written.

Open source. Self-host the moment your team is ready.