RAVIM

01 Product

Persistent engineering intelligence for AI coding agents

Alpha K gives Claude, ChatGPT, Codex and other MCP-compatible coding agents fast, persistent understanding of your repositories — without handing your codebase to another cloud platform.

  • Local-first
  • MCP ready
  • Persistent context
  • Windows · macOS · Linux
Request early access See how it works

Alpha K 1.0.0 · private beta · one-month machine-bound developer licence

02 The problem

Your coding agent should not rediscover the repository on every prompt

Every new conversation starts from zero. The agent lists directories, greps for names and reads whole files to relearn what the codebase is — and pays for every byte of it in tokens, latency and context it no longer has for the actual task.

Measured against Alpha K’s own codebase, a single “analyse the current solution” question cost 151,200 tokens across 12 calls. Eighty-five per cent of that — 128,657 tokens — came from three file reads that pulled 21 whole files at roughly 9.5K tokens each. The next conversation would have repeated it.

03 How it fits

The local control plane for AI coding agents

Alpha K is not an IDE and not a model. Your agent still does the reasoning; Alpha K gives it accurate repository context and bounded access to engineering operations. Pick a question and watch it route.

  1. Agent asksWhere is the token guard implemented?
  2. MCP callworkspace_find
  3. Alpha K routescode graphfile namestext
  4. Bounded resultRanked hits, grouped per file.

Tool names, routing order and response fields are Alpha K’s own. The returned lines are illustrative — run it against your repository to see real ones.

04 Workspaces

Every repository, an explicit workspace

Repositories are added as named workspaces and stay isolated from one another. Each carries its own indexes, its own access rule and its own health.

  • Add and manage repositories as separate workspaces
  • Grant read-only or read/write access per workspace
  • See index status and workspace health at a glance
  • Rebuild intelligence when a workspace needs it

05 Intelligence

Find the right code before spending tokens reasoning about it

01 Code graph Definitions, calls and references traced across the workspace, so an agent can ask what a symbol touches and get a structured answer instead of inferring it from grep output.
02 BM25 retrieval Repository-aware lexical ranking over file names and contents, scoped per workspace, so a query returns the few files that matter rather than every textual match.
03 Workspace briefing A cached orientation to each repository — technology, structure, important files — served before deeper exploration, so a new conversation does not rediscover the codebase.
04 Exact, ranged reads Files are returned line-numbered and byte-exact. An agent asks for the lines it needs rather than pulling whole files into the context window.
05 Impact and references Dedicated tools for what depends on this and what breaks if it changes, answered from the graph rather than reconstructed by the model on every prompt.
06 Published freshness Every artifact reports its convergence state — CURRENT, CATCHING_UP, DEGRADED, BLOCKED or MISSING — so an agent knows whether its intelligence can be trusted right now.

06 The product

This is Alpha K

A desktop Manager on Windows and macOS, alongside a locally hosted MCP server. The rail carries Overview, Workspaces, Intelligence, Operations, Source Control and Settings.

Alpha K Manager Overview: workspace health for two repositories showing code graph, BM25 and briefing all current, above a live feed of MCP tool calls with per-call token counts and timings
Actual Alpha K Manager interface — the Overview surface.

07 Operations

Know what Alpha K is doing

Every tool call is recorded with what it returned and what it cost. The server writes structured rolling logs through the standard logging pipeline, with secrets redacted at the sink, and the Manager reads them whoever started the server.

  • A live feed of tool calls, with per-call token counts and timings
  • Indexing activity and per-workspace convergence state
  • Server status and diagnostics
  • Engineering task execution, visible rather than implicit

08 Source control

Repository actions without unrestricted shell access

Alpha K exposes specific, parameterised operations rather than handing an agent a terminal. Push, merge, deletion, file moves, arbitrary command execution and database mutation are excluded at the tool boundary — not discouraged, absent.

01 Inspect status, diff and log — read the repository state without touching it.
02 Stage explicitly Named paths are staged deliberately. No implicit add of everything that changed.
03 Commit explicitly Commits are an explicit call with explicit parameters, not a side effect.
04 Branch and restore Branch operations and controlled restore of specified paths.
05 Bounded by policy Operations run through a policy layer and input guard, with approval gates on sweeping changes.

09 Local-first

Your source code stays where your source code belongs

Alpha K runs on your machine as a Windows service or a launchd agent. Workspaces point at local repositories and the intelligence is built locally — the repository is not uploaded somewhere else in order to be understood.

  • Repository indexes are built and held on the machine that holds the code
  • Access is explicit per workspace: read-only or read/write
  • MCP provides a bounded tool interface rather than filesystem or shell access
  • API keys are encrypted with OS data protection — never written to plain config, never logged
  • Secrets are redacted at the logging sink, so the log files fail closed

10 Token efficiency

Spend tokens on reasoning, not on rediscovering your repository

A whole-file read above the threshold comes back as a head plus a structural outline. Outlines are produced by a deterministic scanner — no model — so every line is a real line at a real number that can be re-read exactly.

PlatformPaths.cs 1,240 lines · whole-file read

1,240 lines delivered whole file

The whole file enters the context window, whether or not the answer needed it. Above the 400-line threshold, the guard turns this into a head plus an outline.

Any request narrower than the whole file is returned byte-exact, so editing workflows are untouched. Savings are counted in the same encoder used to measure responses, from the copy already in memory. Where nothing was measured, the saving is reported as zero — Alpha K does not estimate what it did not observe.

11 Platforms

Where Alpha K runs

A locally hosted server plus a desktop Manager, installed from one package on each platform. No installer requires a .NET runtime to be present.

  • Windows 10 / 11 · x64
    • Self-contained MSI
    • Server runs as a Windows service
    • Starts before sign-in
    Supported
  • macOS 14+ · Apple silicon
    • Signed .pkg
    • Server runs as a launchd agent
    • Same Manager, same tools
    Supported
  • Linux amd64 · arm64
    • .deb package
    • Server runs as a systemd unit
    • Bound to 127.0.0.1
    Supported

Alpha K speaks the Model Context Protocol, so it works alongside MCP-compatible clients — Claude, ChatGPT, Codex and others — rather than replacing the agent you already use.

12 Licensing

Designed to work when licensing cannot depend on the cloud

Alpha K currently issues machine-bound offline developer licences. There is no licensing server in the path: licences are cryptographically signed and validated locally, and entitlement is enforced at the MCP tool boundary.

  • Install Alpha K and copy the device activation request
  • Receive a signed licence
  • Import it — Alpha K activates locally, offline

Developer licences run for one calendar month.

13 Get Alpha K

Give your coding agent a memory for your codebase

Alpha K is in private beta while we work with a small number of teams on real estates. Tell us what you are building and what your agents are working on.

Request early access

One-month developer licence · Windows, macOS and Linux