Local MCP server + hooks

Stop re-explaining your codebase.

Hologram learns which files, decisions, and prior context your AI assistant needs — before you ask — and injects them into every prompt automatically. It forgets nothing between sessions, and gets sharper the more you work.

Buy Hologram Pro See how it works $99 once · or $12/mo

[HOLOGRAM] (cap=8500B, schema=t3-ctx/v1, project=your-api) tier_counts: hot=3 warm=5 cold=214 - src/auth/tokens.py [HOT 0.94] - src/auth/middleware.py [HOT 0.87] - tests/test_token_refresh.py [HOT 0.81] - src/models/session.py [WARM 0.44] - docs/auth-flow.md [WARM 0.38] … injected before your model reads the prompt. Every session it gets sharper.

actual injection format, routed per prompt

How it works

Learns your project, routes the right context, improves as you work.

Hologram learns how your project fits together, routes the right files and memories into your assistant's context, and gets sharper the more you use it. Under the hood it's a pressure-based router — files heat up when they're relevant, cool when they're not, and every session's usage reinforces what it surfaces next.

Inject

Auto-context on every prompt

With Claude Code, hooks inject routed file context before the model starts — no tool calls, no asking. Gemini CLI, Codex, and Cursor get the same routing as MCP tools.

Learn

A loop that actually closes

Every session's file usage feeds back into the router — including the T3v4-R neural reranker (~1M parameters, trained online on your real usage) and an idle-time consolidation engine.

Own

Local-first, verifiably

Your code, prompts, and routing data never leave your machine. The only network request is license validation. No telemetry. Check it with strace — we publish the manifest.

Who it's for

Built for you if

You work in complex codebases with an AI assistant — Claude Code, Cursor, Cline, Windsurf, Zed, or any MCP-compatible workflow — and you're tired of re-explaining your project every session.

Not for you if

You only work in tiny repos, you don't use MCP-compatible tools, or you want a fully hosted SaaS. Hologram runs locally, on your machine, on real projects.

Trust

The label is machine-verified.

Every path this software reads or writes and every request it makes is published in FOOTPRINT.md — and enforced by an audit-hook test suite that runs a full routing cycle in a sandbox and fails if anything outside the manifest is touched, or if any socket opens beyond license validation. Install and uninstall print every change before making it, and uninstall reverses all of them.

If your license ever lapses, Hologram degrades to its open-source core — pressure routing keeps working. It never breaks your setup, and it never holds your workflow hostage.

Revocation policy: degrade, never brick

Quickstart

Five minutes to routed context.

After purchase, your license key and the installer wheel are both waiting on your Polar portal page. Then it's four commands:

# 1 — install the wheel you downloaded from your portal
pip install './hologram_cognitive_pro-*.whl[mcp,hooks]'

# 2 — activate with your key (also on your portal page)
hologram license activate HP-XXXX-XXXX

# 3 — hook up your editor (prints every change first)
hologram install claude-code

# 4 — first route
cd your-project && hologram init .claude

Pricing

Two ways to buy. Same product.

Perpetual

$99 once

Perpetual license, 3 machines
1 year of updates included
All pro features forever

Buy perpetual

Subscription

$12 / month

3 machines, cancel anytime
Updates while active
Lapses degrade to open core — never breaks

Checkout is handled by Polar (merchant of record — VAT and invoices included). Your license key appears in your purchase page immediately. The open-source core, hologram-cognitive, is MIT-licensed and free forever.