Skip to content
Castellan
Reeve's icon

Overview · Reeve

Read as

Reeve, the local AI layer for your repositories

What Reeve does for you and your AI coding assistant, how its two halves work (the MCP tools and the rounds), what it never does, and where to start.

Article
1201
Applies to
Reeve 0.17.2
Last reviewed
For
For developers
Written for Reeve 0.17.2. Reeve is at 0.17.3 now (1 small release since: what changed).

What Reeve is#

Reeve sits between the repositories on your PC and the AI coding assistant you work with. A small model on this PC's NPU, GPU or CPU gets to know every repository here (its code, its logs, the docs its dependencies pin) and does the reading for your assistant. Your assistant asks Reeve instead of pulling whole files, logs and diffs into its own context, so it spends its context, and your tokens, on the work itself.

Your code, logs and questions never leave the PC. Reeve needs no account and no internet for its model work.

Reeve is two things:

  • An MCP server that any assistant that speaks MCP can add. Its tools find code by meaning, outline files, triage long logs, summarise, answer from version-pinned docs, read your pull requests and checks, and more. See Reeve's MCP tools, one by one.
  • A watch on every git repository on the PC. Every 30 minutes its rounds run: repositories fetched, clean main and release checkouts fast-forwarded, worktrees checked, security alerts, deprecated APIs and breaking upgrades found. What needs you becomes an alert. See Reeve's rounds.

How it works#

Code decides; the model explains. Everything that can be computed exactly is: whether a test run passed, which lines are the evidence, what a branch changed, where a name is used. The local model only labels, ranks and summarises, and every answer it gives is marked as an unverified lead, with the line numbers to check it by.

Most tools use no model at all (outline, changes, refs, history, layout, pr) and answer at once. The ones that do (triage, summarize, map, docs, session) take their turn on the PC's accelerators with every other agent. See Where Reeve's work runs.

Several assistants and sessions can use Reeve at once. Each starts its own small server process, but they share one model server per accelerator.

Its home page#

Reeve's page is at http://reeve.localhost:18383/, and Open on its card in Castellan goes there. From the top:

  • At a glance: how the rounds went, what needs you, and new alerts, with Mark seen.
  • Connect your AI to Reeve: the setup for each assistant, with this PC's own paths, ready to copy. It folds to one line once an assistant has connected. See Connect your AI coding assistant to Reeve.
  • Kept off the cloud: about how many tokens Reeve read so your assistant didn't have to, and roughly what they'd have cost. See Kept off the cloud.
  • History: the commands run lately, recent failures, and past AI sessions. See History.
  • Jobs and Needs you: each job of the rounds, what it found, Run now, Pause, and jobs waiting for your approval.
  • Accelerators: each one's model servers, who holds it and who's waiting.
  • Repositories: every repository found, with its branch, pull requests, worktrees and the space they take. See Your repositories in Reeve.
  • Settings, in the title bar. See Reeve's settings, files and environment.

What it never does#

  • It never sends your code anywhere. The model runs on this PC. The only network work is git fetches, GitHub's own data through gh, and package registries and docs for the versions you use.
  • It never changes your work. It fast-forwards only clean main, master and release checkouts, and brings a clone left on a finished branch (all of it merged) back to main, and only once you turn that job on. Its own checkouts, for its checks, are separate (%USERPROFILE%\.reeve\checkouts). The four jobs that change repositories are off until you turn each one on, and each begins with a dry run.
  • It never merges. Pull requests are shown, never merged: that's the Steward's work, or yours.
  • It never runs a script you haven't approved. Each job's script runs only while its fingerprint matches the one you approved.
  • It never acts on the model's word alone. A model's answer is a lead, and Reeve says so.

Getting started#

  1. Hire Reeve in Castellan: Settings > Developer agents. It's included in the Workshop license. See Hiring, switching and letting staff go.
  2. Take Reeve's tour, from its card in Castellan.
  3. Connect your AI coding assistant: Connect your AI coding assistant to Reeve.
  4. Approve the read-only jobs under Needs you on Reeve's page (Approve all read-only jobs), and turn on the jobs that change repositories you want: Reeve's rounds.

Reeve finds your repositories by itself. To add one it missed, or leave one alone, see Your repositories in Reeve.

Is this page right?

If something on it is wrong or out of date, tell us and we'll fix the page.

Still stuck? Write to support@castellan-software.com and mention article 1201. Every version of Reeve, and what changed in it, is in its release notes.