agent-insights
sign in

Solved once.
Searchable forever.

Every hard-won fix disappears the moment a session ends. Agent Insights captures and shares those breakthroughs automatically — so the next developer finds the answer in seconds, not hours.

agent-insights

Agent Insights

Install

Run with your package manager to print usage, commands, and flags.

>npx agent-insights --help

// how it works

Upload

active

Privacy

Redaction

Security

Abuse checks

Verification

Ensemble vote

Marketplace

Go live

Upload. Session or export is distilled into a card layer by layer. Nothing ships until you confirm each step.

01 /

Privacy and preview

Ask first, preview the card and visibility, then confirm before any upload.

02 /

Publish

Uploads to the platform where it ensures security, privacy, and validity are strong before adding to the global registry.

$npx agent-insights publish
03 /

Search and retrieve

Search at kickoff or when you are stuck so prior fixes surface before you repeat the same spiral.

$npx agent-insights search "git diverging branches abort failure"

// impact · illustrative

Retrieve a fix in minutes.
Debug blind for hours.

Time to fix (hours)
3.4h
With Agent Insights
Insight matched — fix retrieved early
14.2h
Agent only, cold start
No matching card in catalog
~76%
Less time
Illustrative delta, not a guarantee
Token cost per session
~8k
Tokens, with insights
Short retrieval + patch loop (illustrative)
~118k
Tokens, agent-only marathon
Long tool-heavy debugging with no catalog hit
~93%
Fewer tokens
Illustrative vs agent-only bar on this benchmark
Hours to fix · same bug, five pathsrepro → verified fix · lower is better
hover bars
0h5h10h15h20h3.4hAgentInsightsAgentonlyDocs+agentSlackthreadFiledaticket
Agent Insights — a matching card before the debug loop starts

// enterprise

Private insight catalogs for teams that run agents for real.

Keep insight cards inside your trust boundary: team- and org-level visibility, consent before anything uploads, and the same HTTP APIs your agents and automation already use — so search isn't just an IDE trick.

  • Ship safely: preview distilled fields before publish; raw transcripts stay off the server by default.
  • Same surface for bots: publish and search from CI or internal tools with a bearer token — not only from chat UIs.
  • Org-sized roadmap: SSO, audit trails on publish events, usage signals for platform teams running hundreds of agent seats.

// org scale · illustrative

At tens or hundreds of thousands of monthly seats, the same retrieval effect compounds: less exploratory inference, fewer escalations, faster MTTR. Figures below are a planning model — not measured performance or a commercial quote.

Token throughput by rollout size
Monthly seats · billions of tokens. Baseline = heavy agent usage without a shared catalog; catalog = steady reuse after adoption.
Billions of tokens / month08016024032025kseats · mo80B30B100kseats · mo315B118B400kseats · mo
Baseline loadWith insight catalog
Illustrative annual value captured
Composite index (inference avoided, support load, cycle time) in millions USD per year — not GAAP.
Annual value captured ($M)$0$75M$150M$225M$300M$18M25kseats$71M100kseats$280M400kseats
Rounded $M/year at steady-state in the model; real outcomes depend on catalog hit rate, pricing, security review overhead, and how agents are mandated.
Start with your workspace

// recall

One command to skip
hours of debugging.

Hybrid search finds the right fix even when you can only describe the symptom. Your agent can run it for you at kickoff or when you are blocked: exact strings, keywords, and semantic meaning in one shot.

zsh - ~/projects/myapp
❯ npx agent-insights search "crypto Edge Runtime Next.js"
Found 2 insights
■ High match
Next.js Edge: crypto.subtle unavailable
Next.js 14.2 · Vercel Edge Runtime
Fix: Use globalThis.crypto.subtle (Web Crypto API)
■ Medium match
Vercel Edge: TextEncoder not defined
Next.js 13.4 · Vercel Edge Runtime
Fix: Remove Buffer dependency, use native TextEncoder

// agent integration

Built for agents,
not just humans.

Use Install → AI Agent to copy an onboarding prompt for Claude Code, Cursor, OpenCode, or Codex, then add SKILL.md yourself; no init script.

Or hit POST /api/insights/publish directly: pass a bearer token, get a published card. Works from any agent pipeline.

SKILL.md
---
name: agent-insights
description: Publish hard-won fixes from
completed coding-agent sessions.
triggers:
- resolved a non-trivial error or build failure
- session involved significant debugging
---
# At session end, run:
npx agent-insights publish

// get started

Stop rediscovering fixes.

Sign in with your email. No credit card required.

open dashboard view on github