Skip to content
C-F0x edited this page Oct 26, 2026 · 2 revisions

CompileGateway

CompileGateway pushes a source tree from your machine to an SSH target (usually a macOS build box), runs the build there, and brings the artifacts back or packages them for export.

It is one core implementation with four front ends: a browser UI, a command-line client, a built-in MCP server for AI agents, and the HTTP API they all use.

What it does

  • Push — send a local project directory to the target machine, with sensible excludes and CRLF→LF normalization for scripts.
  • Build — run the target machine's own toolchain through a declarative build type (manifest): variables, ordered steps, artifact paths.
  • Bring results back — pull artifacts to the host, export them as a zip, or copy/move/delete files on either side.

Four front ends, one implementation

Front end What it is Use it when
WebUI browser UI served by the daemon you want to click through projects, jobs, logs and settings
CLI (cgw) thin HTTP client, never prompts scripts, terminals, anything repeatable
MCP built-in server on its own port you want an AI agent to drive the builds
HTTP API the API the other three use you are writing your own client

Parity principle: Whatever the WebUI can do, the CLI can do. The reverse is intentionally scoped (machine-oriented capabilities like raw JSON Schema retrieval and workflow-preset command-line application are CLI/MCP/HTTP only to keep the WebUI clean and focused).

Requirements

Rust (nightly) and Node 18+ to build; WebView2 on Windows for the desktop shell; an SSH target machine; the target's own toolchain (Xcode, Gradle, …) for builds.

Quick start

# 1. build the WebUI first (it is baked into the binary)
cd webui-src && npm ci && npm run build && cd ..

# 2. build the binaries
cd src-tauri && cargo build --release && cd ..

# 3. start the daemon (headless) or the desktop shell
src-tauri/target/release/compilegateway-daemon

# 4. point the CLI at it (it finds the daemon by itself in most cases)
cgw doctor

Then: add a client (SSH target), register a project, push, pick or write a build type, build, and pull or export the result. Getting started walks through it.

Documentation

Page What it covers
Getting started build, run, first project, first build
Concepts the object model, job lifecycle, concurrency rules
Tasks the seven task kinds, input/output contracts, artifacts
Batches and workflows ordered multi-step runs and cross-project templates
Build types (manifests) describing how a kind of project is built
Clients and environment SSH targets, host keys, base dir, environment probes
WebUI the browser interface, page by page
CLI cgw: commands, exit codes, scripting
MCP the built-in AI-agent server
HTTP API endpoints, auth, error codes
Configuration and logs config.json, three ways to change it, log files
Troubleshooting symptoms, error codes, what to try

At a glance

Task kinds build, push, pull, copy, move, delete, export
Concurrency one job per project; per-project FIFO queue with enqueue; optional host-wide lock
MCP 42 tools, 8 resources, 3 prompts, its own port (default 3994)
Schemas manifest, task, batch, client, project, workflow — generated from the Rust structs
State on disk config.json, runtime.json, projects.json, jobs.json, manifest/, workflows/, logs/

Clone this wiki locally