Rust·Latest release v0.3.61 · 2026-11-16

prod-code

Remote code intelligence and compilation daemon for coding agents.

Latest releasev0.3.612026-11-16
Last committoday2026-11-16
LicenseMIT OR Apache-2.0
Articles37on this blog
Recent releases
All 9 releases →
Recent commits
  • 67abfb5 release: prepare v0.3.61 (#2037) 2026-11-16
  • 9477e96 code_outline of a shell script is not routed by its language (#2036) 2026-11-16
  • bbd7d1b code_test of a SwiftPM package is routed to a node with swift; a refused code_source read names what the gateway serves (#2034) 2026-11-16
  • 25ae316 release: prepare v0.3.60 (#2031) 2026-11-14

Status from GitHub as of 2026-11-16.

What it is

prod-code moves code intelligence, builds and tests for coding agents off the laptop and onto a machine on the local network. The analyzer, the builds and the tests run on the node; the laptop edits.

A language server was designed for one human in one checkout. A fleet of agents runs a resident agent in the main checkout and workers in their own worktrees, each asking the analyzer questions many times a second and each bringing its own analyzer and its own test run. prod-code gives every worktree its own warm analyzer on a remote node and exposes it to agents through MCP, a CLI and LSP.

  • Hover on a warm workspace in 1-4 ms; a tool call over a persistent session in ~10 ms.
  • cargo clippy on a 32-core node with one file changed: 1 s.
  • Edits are validated on a second analyzer before they are written, so a dry run never slows anyone else down.
  • Rust, Go, C/C++, TypeScript/JavaScript, Python and Swift through their own language servers.

How it works

The checkout is mirrored, not mounted

First contact sends a manifest of paths, sizes and hashes; the node seeds a worktree and asks only for what is missing (0.4 s instead of ~7 s for a 10 MB repository). After that the client sends diffs, and the MCP server syncs only when files changed.

One analyzer per worktree

Every Git worktree gets its own server workspace and database, so worktrees never answer each other’s questions. An engine stays resident until it has been idle for thirty minutes.

Proposals are checked on a second analyzer

Validating an edit runs on a separate engine for the same workspace, fed by the same syncs, so a dry run never invalidates the analyzer everyone else is querying.

Builds and tests run on the node

Commands run in the workspace’s own warm target directory; files they change are written back into the checkout. Several nodes gossip and place each checkout on the node that already holds it or the quietest one that serves its language.

Example uses

01. Symbol navigation

Navigate a repository by symbol through the configured remote analyzer.

02. Proposed edit checks

Inspect analyzer diagnostics for a proposed edit before applying it.

03. Remote test runs

Execute configured tests on the remote node.

Where it fits

prod-code combines a client with a remote gateway and language toolchains. Each worktree uses isolated analyzers and caches.

Get started

Follow the repository setup for the client, remote gateway, and language toolchains.

Build the client from source with Mitts and install it on PATH (rustup puts ~/.cargo/bin there)

# needs Rust 1.95+ and Mitts: https://prod.codes/projects/mitts/
git clone https://github.com/alex09x/prod-code && cd prod-code
mitts build --release -p prod-code-client --bins
install -m 755 target/release/prod-code "$HOME/.cargo/bin/prod-code"
prod-code --version

Register it once for every repository

claude mcp add --scope user prod-code -e PROD_CODE_REMOTE=<node>:9400 -- prod-code mcp

Run a node

prod-code-server --bind 0.0.0.0:9400 --storage /srv/prod-code/workspaces \
  --advertise <this host>:9400 --peers <other nodes>

README →

Tools an agent gets →

Changelog →

Limitations

  • A reachable configured node and its toolchains are required.
  • Initial indexing takes time.
  • Tool support varies by language.
  • Edit validation returns analyzer diagnostics; it does not cover every compiler or borrow-checker error, so remote compilation and tests remain necessary.
Cite this software · archived version v0.3.49

Cite this software

Alexander Panasenko (2026). prod-code — Remote Code Intelligence for Fleets of Coding Agents (v0.3.49). Zenodo. https://doi.org/10.5281/zenodo.23905176

This record reflects the archived software version deposited on Zenodo for citation and reproducibility, which may differ from subsequent development releases.