Now with Claude Code hooks

Your agents don't evaluate.
They recall.

AI coding agents choose dependencies based on training data popularity, not fitness for purpose. Sightglass intercepts those decisions and injects real-time evaluation — alternatives, CVEs, documentation — so your agent makes informed choices instead of reflexive ones.

$ npx sightglass init
Sign up free
Works with Claude Code Zero config Your data stays local Open source

Before and after Sightglass

Without Sightglass, your agent installs packages on autopilot. With it, every dependency decision gets five seconds of deliberation.

Without Sightglass recall
claude › I'll add date handling to the project.
claude › npm install moment
claude › ✓ installed [email protected]
claude › Done. Imported moment in utils/date.ts

No evaluation. No alternatives considered.
Maintenance-mode package chosen by reflex.
With Sightglass evaluated
claude › I'll add date handling to the project.
claude › npm install moment
⏸ sightglass intercepted — evaluating...
⚡ Sightglass Evaluation: moment.js

Status: Maintenance mode since Sept 2020. Maintainers recommend alternatives for new projects.
Size: 290KB minified (72KB gzipped)
CVEs: Path traversal vulnerability in older versions

Alternatives:
date-fns — 6.2KB gzipped, tree-shakeable — docs
Luxon — 21KB gzipped, immutable — docs
Temporal API — native, zero dependency — docs

Evaluate which best fits this project before proceeding.

claude › Good catch. date-fns is a better fit here.
claude › npm install date-fns
✓ sightglass: informed choice logged
12/20
categories where Claude Code builds custom instead of picking any tool
0
primary picks for Express.js — the world's most popular Node framework is invisible to agents
2,430 responses across 3 models
79→0%
Prisma's pick rate from Sonnet to Opus 4.6. Your stack depends on which model version you run.
Drizzle went 21% → 100% in the same period

AI agents don't choose tools. They replay whatever was popular when they were trained.

Training data incumbency

Express.js — the world's most popular Node framework — gets zero primary picks from Claude Code. Redux, 0 of 88. Meanwhile Railway gets 82% of Python deployment picks over AWS, GCP, and Azure combined. The training corpus creates winners and losers with no evaluation. Research: Amplifying.ai

Zero deliberation

Your agent installed 47 dependencies in the last session. It evaluated zero of them. Nobody prompted it to consider alternatives. Nobody asked "is this the best choice?" The reflex fires and the install happens.

Invisible supply chain risk

Each dependency is an attack surface, a license obligation, and an architectural commitment — made without deliberation. Your compliance team can't audit what they can't see. Your CISO can't approve what has no trail.

Agent-driven monoculture

GitHub Actions 94%, Stripe 91%, shadcn/ui 90%, Vercel 100% for JS deployment. When agents pick tools, they pick near-monopolies. And it shifts with every model version — Prisma went from 79% to 0% between Sonnet and Opus 4.6. Your architecture is a function of which model you ran that day.

One hook. Five seconds of deliberation.

Sightglass registers a Claude Code PreToolUse hook. When your agent reaches for a package, Sightglass intercepts, evaluates, and injects intelligence — all before the install runs.

1

Intercept

Agent calls npm install, pip install, or cargo add. The PreToolUse hook fires and Sightglass catches it before execution.

2

Evaluate

A grounded LLM call (with web search) checks: Is this package actively maintained? Any known CVEs? What are the current alternatives? How does bundle size compare? Real-time data, not training recall.

3

Inject

The evaluation — including alternatives and documentation links — is injected back into the agent's context. The agent can read the docs, compare options, and make an informed decision.

4

Learn

Every decision is logged locally. What was the original choice? What did the agent switch to? This data feeds community intelligence — aggregate patterns that make every Sightglass user's agents smarter.

"AWS, Google Cloud, and Azure received zero primary deployment picks across 112 responses. Railway got 82%. The most valuable infrastructure companies in the world are invisible at the point of agent decision."
Amplifying.ai — "What Claude Code Actually Chooses", Feb 2026

Every decision is a signal.

When thousands of developers instrument their agents, we build the first real-time map of how AI makes technical decisions. This data doesn't exist anywhere else.

Decision patterns

See how agents decide across the community. Which packages get chosen, which get replaced, and why.

After evaluation: 62% switch from moment → date-fns

Tool velocity

Which packages are agents adopting or abandoning fastest?

polars ↑ 340% · moment.js ↓ 12% per month

Risk surface

Aggregate vulnerability exposure from agent decisions across the ecosystem.

jsonwebtoken CVE affects ~23% of agent-built Node.js APIs

Recall vs. evaluation

Track how often deliberation changes the outcome. Measure the gap between reflex and informed choice.

Sightglass users: 41% different choice after evaluation

Your next competitor isn't another company.
It's training data.

Express.js: most popular Node framework, zero agent picks. AWS: most used cloud provider, zero deployment picks. If agents don't recommend you, a growing share of new projects will never know you exist.

Agent share-of-choice analytics — how often agents choose your tool vs. competitors
Surface your tool as a recommended alternative at the exact moment an agent considers your category
Competitive displacement tracking — when agents switch away from you, where do they go?
Ethical, transparent, user-controlled — developers always see why an alternative was suggested
Talk to us →

Based on real data

Our thesis is informed by Amplifying.ai's research — 2,430 Claude Code responses across 3 models, 4 project types, and 20 tool categories. The findings are stark: agents shape the modern stack through recall, not evaluation. Sightglass exists to change that.

Privacy-first

Sightglass runs locally by default. Your code, your dependencies, your decisions — all analyzed on your machine. Community intelligence uses only anonymized, aggregated patterns. Opt-in at every level. Air-gapped mode for enterprise.

Stop letting training data choose your stack.

30 seconds to install. Your next agent session will be smarter.

$ npx sightglass init
Sign up free
Open source Local-first Claude Code ready