For people who work with AI agents

Reeve and Workshop

Your AI coding assistant is the brain. Reeve, a local MCP server for your repositories, does the reading.

Reeve

The local AI layer

Your AI coding assistant is the brain. Reeve does the reading. It's a small model on your own NPU, GPU or CPU that learns every repository on your PC: the code, the logs, and the docs each project pins. Your assistant asks Reeve instead of loading whole files and logs into its own context, and gets back the few lines that matter.

Fewer tokens, smaller bills, limits that last longer. The bulk reading happens on your hardware, for free.
Bounded, checkable, high-volume work. Code decides what can be decided (did the test pass?); the model explains. Every model answer is labelled for your assistant to check.
Private and offline. Your code, logs and questions stay on the PC.

Reeve

A local MCP server

Reeve speaks the Model Context Protocol, the standard way AI coding assistants call tools. Its MCP server is a process your assistant starts on this PC and talks to over standard input and output (stdio): not a web service, nothing listening on the network.

When it connects, Reeve hands your assistant its own instructions: which tool fits which job. Your assistant needs no prompt of its own to use it well.

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

Reeve

Connect your AI coding assistant

When you hire Reeve, it registers itself with the most common AI coding assistant's command line if that's installed. Its page, Connect your AI to Reeve, has ready-to-copy setup, with this PC's paths filled in, for the popular assistants and editors.

For any other MCP client, add a stdio server that runs Reeve's command:

MCP client settings (JSON)

{
  "mcpServers": {
    "reeve": {
      "command": "node",
      "args": ["C:\\Users\\<you>\\.reeve\\app\\src\\cli.ts", "mcp"]
    }
  }
}

Then ask your assistant something Reeve is good at, like "why did the last test run fail?" or "where do we validate the licence key?", and watch it call triage or search.

Reeve

The tools

Fifteen tools, each a job an assistant would otherwise do by reading everything itself:

triage: Digests long build, test and runtime logs. Code decides pass or fail; the model labels the causes. For a failed UI test it reads the screenshot too.
search: “Where is X handled?” by meaning, across a repository, from an incremental index. Under a second once indexed.
summarize: The gist of a big file, with line references back to it.
map: The same one-line question over many files or items at once.
docs: Answers from the docs stored for the exact versions your lockfile pins, with every quote checked against them.
docs_list: Which docs are stored, and the project's dependencies with their versions.
lint_docs: Code that uses APIs your dependencies mark deprecated, removed or renamed.
upgrades: The breaking changes in outdated dependencies that hit APIs this code uses, with the file and line.
screens: Clipped, overlapping or cut-off text in your app's screenshots, found with Windows' own OCR.
cluster: Groups near-duplicate texts: crash logs, test failures, feedback.
ask: One small, low-stakes job, like drafting a commit message.
digest: The latest result of each of Reeve's rounds, and the alerts not yet acknowledged.
repos: Every git repository on the PC, with its branch, lockfiles and what the rounds do there.
manor: Castellan's whole state in one call: alarms, findings, agents down, updates, accelerators.
status: The accelerators, models and index, and the tokens kept off the cloud.

Reeve

Where the AI runs

Reeve runs small open models: one for text, one that can see screenshots, and one for search. It picks the best accelerator on your PC:

The NPU first, whoever made it: Snapdragon, Intel Core Ultra, or AMD Ryzen AI 300 and later. Fastest, coolest and lightest on the battery.
Then the graphics card.
Then the processor, so it works on any PC.

The Smith keeps the model servers running for every agent, and stops the ones nobody is using so their memory comes back.

Reeve

Tokens kept off the cloud

Reeve's page keeps a running count of the tokens it read so your assistant didn't have to: this month, last month and all time, with what they would have cost.

Kept is what Reeve read minus what it handed back. Only text from your own disk counts, and the same input read twice within an hour counts once.
One input counts at most 25,000 tokens and one call at most 100,000, so a giant log doesn't inflate it.
The price is $3 per million tokens unless you set your own. They're estimates, and the page says so.

Your repositories

Rounds every 30 minutes

Every 30 minutes Reeve looks after every git repository on your PC and leaves a digest on its page, so your assistant (and you) start the day knowing what changed:

Fetches, and fast-forwards clean main, master and release/* branches. It never touches a branch you're working on.
Git hygiene, security alerts, deprecated APIs and the breaking upgrades that would hit your code.
Triage of test runs, so a failure arrives with its cause.

Reeve watches. Your assistant reads and acts. You acknowledge.

Your repositories

Which repositories

Every git repository Reeve finds in the usual folders (Projects, src, repos and the like, on every drive), every folder your AI coding assistant has worked in, and any you add.

Leave alone on its page takes a repository off the rounds entirely.
The search index lives in your user folder, never inside the repository, and leaves out what .gitignore leaves out.
A project with a lockfile gets its own separate checkout for Reeve's checks, its dependencies' docs at the pinned versions, and dependency health. Packages are installed with their install scripts off.

Workshop

The developer roles

Reeve is one of Workshop's developer agents. Workshop has everything in Household too, and its developer agents take their offices while the household's keep working behind them.

Reeve: The local AI layer under your AI coding assistant: a local MCP server for your repositories.
Developer Herald (coming soon): News from the projects your code is built on.
Aletaster (coming soon): Tastes each release of your own projects before it goes out.
Pinder: Finds stray processes and the ports they hold.
Steward (coming soon): Keeps your own repositories in order.

Turn on Developer options in Castellan's Settings (or pick "Writing software" when you first set it up) and the developer roles open.

Questions

Does my code leave my PC?: Not through Reeve. Its model, its index and the docs it stores are all on your PC. Your assistant decides what it sends to its own service; with Reeve, that's the short answer instead of the whole log.
Which assistants work with it?: Any that can use an MCP server over stdio: the popular coding assistants, editors and command lines all can.
Does it replace my assistant?: No. It does the reading so your assistant can do the thinking, with less to read and fewer tokens spent.
Do I need an NPU?: No. It runs on your NPU, your GPU or your CPU, best on an NPU.
What does it cost?: Reeve is in Workshop: $8 a month or $72 a year, for up to 3 PCs, with every agent in Household too. 7 days free, no card.