The living encyclopedia of AI coding

Improve how you build with AI. Automatically.

Your agent knows your code and workflow. Dorsal shows it what 48,420 other builders learned. Updated daily.

Start with a 3-minute audit of your recent work.

Find three things to improve now.

Terminal · Connect via MCP

claude mcp add-json dorsal '{"type":"http","url":"https://godorsal.com/api/truthapi","headers":{"Authorization":"Bearer tk_alpha_demo_001"}}' --scope user

Claude Code · Run this

dorsal audit

Read-only · Raw history stays local

“Dorsal improved what I do today, in minutes. Now I want more.”
Boris Pavlovic · Founder & CEO, itlook.ai

Remove hidden blockers

Catch upstream bugs and silent failures your repo cannot show you.

codex

›

dorsal audit

•

I’ll inspect the recent work locally, then check every proposed improvement against Dorsal MCP. Raw history stays local.

61 Dorsal MCP calls

10 concurrent sub-agents · 2 adversarial rounds

~120 distinct sources inspected

•

7 of 21 proposed fixes would fail or backfire. 5 would fail silently.

Do not replace prose rules with hooks

Hooks do not reliably fire for sub-agent Bash. Removing the rules would leave no gate.

Do not add the longer HALT clause

The goal text is resent until the evaluator jams. The proposed fix triggers a worse version of the bug.

Do not trust the git-status digest

Re-editing an already-dirty file does not change it. The receipt can certify untested code.

Make good features awesome

Find the missing behavior, sharp edges, and workflow friction between “works” and “I use this every day.”

codex

›

dorsal improve this homepage agent. Find what separates useful product demos from generic chatbots.

•

I’ll compare measured successes, failed first-use experiences, and the MCP servers developers keep using.

48 Dorsal MCP calls

38 getSummaries · 5 related360 · 4 getSet · 1 getWaves

• Called mcp__dorsal__getSummaries

{
  "query": "homepage embedded AI chat agent interactive demo converts visitors into users activation adoption conversion",
  "kind": [
    "ux",
    "agent",
    "app",
    "service"
  ],
  "total": 40
}

83 records

• Called mcp__dorsal__getSummaries

{
  "query": "homepage AI chat agent onboarding failure empty prompt trust latency citations user experience",
  "kind": [
    "ux",
    "limitation",
    "reliability"
  ],
  "total": 40
}

80 records

•

The winning pattern is not a generic chatbot. It is interactive proof of the product’s core value.

Perform the product’s job

Let the visitor make one consequential technical decision before signup.

Show what changed

–Lead with the recommendation
–Expose the evidence that changed it
–Show the strongest counterevidence

Close the loop

End with one concrete change to the plan. Offer MCP installation only after delivering value.

Result

Rebuilt the homepage agent around this pattern. “Dramatically better. Night and day.”

Evaluate ideas against prior art

See who tried it, what failed, and whether the gap is real—in seconds, not hours.

codex

›

dorsal evaluate this coding-history audit idea against prior art and known failure modes.

•

I’ll test the idea against existing retrospective products, current research, privacy failures, and false-audit reports.

17 Dorsal MCP calls

11 getSummaries · 5 related360 · 1 getSet

~1,410 summary records · 16 full records inspected

• Called mcp__dorsal__getSummaries

{
  "query": "Claude Code Codex coding history session transcript audit analyze past work mistakes recurring failures personalized best practices"
}

retrospective products and practices ranked

• Called mcp__dorsal__getSummaries

{
  "query": "session retrospective false conclusions overgeneralizing from one failure human approval repeated pattern threshold privacy local processing"
}

failure modes and safeguards ranked

• Called mcp__dorsal__getSet

{
  "ids": [
    "gh-anthropics-claude-code-issues-78991_issue_filed",
    "gh-obra-superpowers-issues-864_issue_filed",
    "gh-openai-codex-issues-31588_issue_filed"
  ]
}

3 of 16 decisive records shown

•

Strong first-use hook: 9/10. Change the promise before building it.

Demand exists

Dreamer collapsed 235 feedback instances into 11 recurring issues. Other tools already mine agent history.

Do not promise “everything wrong”

Claude Code /insights has mistaken its own pipeline failures for user failures and cached the false diagnosis.

Build

Keep raw history local. Require recurrence across independent tasks. Return three evidence-backed improvements and ask before changing anything.

Choose the right design

Compare architectures against your constraints. Pick one. Record why.

codex

›

dorsal choose the right cross-client architecture for this audit feature. Compare Codex-first, Claude-first and one local core.

•

I’ll compare Codex-first, Claude-first, dual-agent, and agent-native designs against current client behavior and failure reports.

12 Dorsal MCP calls · 2 web searches

Parallel 30–40 record scans · 13 full records inspected

• Called mcp__dorsal__getSet

{
  "ids": [
    "gh-anthropics-claude-code-issues-76362_issue_filed",
    "gh-anthropics-claude-code-issues-78991_issue_filed",
    "gh-obra-superpowers-issues-864_issue_filed"
  ]
}

3 of 13 decisive records shown

•

Build one local core. Dogfood Codex first. Add Claude Code before the external pilot.

Why Codex first

Its local thread store, rollout JSONL, skills, plugins, hooks, and configuration are inspectable.

Why one local core

Parsing, redaction, task reconstruction, evidence adjudication, and reporting should not drift by client.

Do not trust hooks as history

Hooks miss print-mode exits, active transcripts, and work before compaction. Use them as triggers, not evidence.

Find the best tech

Rank libraries and services by fit, integration cost, failure modes, and current docs.

codex

›

dorsal find the best authentication stack for Xapa’s persistent user sessions.

•

I’ll rank the stack against Xapa’s Flutter clients, enterprise SSO, daily-return loop, deep links, and account lifecycle.

Initial Dorsal MCP calls completed in parallel

240 records per getSummaries query · 8 WorkOS connections

5 decisive full records inspected · 1.6s

• Called mcp__dorsal__getSummaries

{
  "query": "web application authentication architecture current best practices OAuth OIDC passkeys WebAuthn magic links sessions cookies CSRF account linking multi-tenant organizations RBAC security",
  "total": 100
}

1,120 ms

• Called mcp__dorsal__getSummaries

{
  "query": "best authentication providers libraries for product-led SaaS Next.js React TypeScript Clerk Better Auth Auth.js Supabase Auth WorkOS AuthKit Stytch Descope Auth0 comparison implementation",
  "total": 100
}

1,476 ms

• Called mcp__dorsal__related360

{
  "query": "WorkOS AuthKit",
  "hops": 2,
  "direction": "both",
  "total": 100
}

8 connected records · 511 ms

•

Evaluate Auth0 as the unified option against the existing session system plus WorkOS.

Product decision

Authenticate before entry. Preserve the destination Quest. Make authentication invisible during the session.

Required architecture

Rotating mobile refresh tokens, secure web cookies, silent renewal, server-side membership checks, SSO, and SCIM.

Not the default

Better Auth. Xapa’s Flutter and enterprise lifecycle requirements materially change the ranking.

dorsal <anything>

Probe any decision, claim, design, technology or problem against what builders already learned.

dorsal research thisdorsal examine thisdorsal improve thisdorsal audit

Your agent interprets the context. Your code stays local.

Or call getSummaries · getSet · related360 directly.

Ask Dorsal about a real decision.

One complete, source-backed answer. No signup.

48,420 builders. 11,436 products. 9,211 methods. 4,284 problems. 1,096 use cases. Connected to the sources that show what happened.

Explore the data

Adversarial audit

7/7

proposed fixes would fail

would fail silently

The agent proposed the fixes. Dorsal found the failures.

Claude Code's own extensive audit produced seven unsafe fixes. Dorsal checked them against upstream bugs, release changes, and failures reported by other builders before any code changed.

Inspect the example

Why Dorsal catches what your agent misses

The current technical record. Connected.

Official documentation, releases, GitHub issues and PRs, research, incidents, and builder reports—linked across products, methods, problems, and use cases.

48,420

builders

14,234

products

14,821

methods

6,157

problems

2,491

use cases

74,940

canonical documentation chunks

Every finding retains its source, date, authority, and lifecycle—including what is superseded or contested.

“Dorsal is integrated into the Keelo Brain and used every day to discover technologies, methods, and open-source components for specialized customer agents.”

Founder, Keelo.ai