The grounding, verification & orchestration layer for AI work.

Turn your team's rules, facts & standards into the rails your AI runs on.

Booster grounds every agent before it acts. Inspector checks every output after — and when work leaves the rails, it returns a pass/fail verdict with a machine-readable rejection code and a cited, auditable trail. And Nora runs the whole loop — ground → act → verify — with as much or as little human in it as your verification coverage earns. Portable across Claude, Cursor, Copilot. Model-neutral. Yours to own.

  • 01
    Ground before it writes Booster supplies the facts that govern the work — the agent reasons from your truth, not the internet's average.
  • 02
    Verify after it writes Inspector checks every output against those same facts and returns a cited pass/fail verdict — with the rule it broke.
  • 03
    Open Factlet protocol Git-native, vendor-neutral, W3C PROV-O lineage. Not a proprietary format.
How it works — the Nora loop
Nora at a glance
LOOPS BACK THE STORE Factbook curated · git-native EVERY STAGE READS & WRITES 1 Capture extract decisions → factlets 2 Ground injected before the AI writes 3 Inspect verified · cited audit trail 4 Maintain supersedes · retires stale INPUTS source repos docs policies AI sessions SURFACES Desktop Dashboard CLI MCP Server IDE Panel Agent SDK Agents AWS Managed Claude Code VS Code Kiro Cursor chat Report + audit trail verified · cited trace
Nora in Claude Code
Engineering
LYou · 11:42

How should I handle Stripe webhook retries here?

NClaude · grounded by Nora

For this codebase, use an idempotency key on every handler f024 and exponential backoff with jitter, capped at five retries f117.

Don't retry without dedup — that caused incident f214 last quarter (double-charge on /refund).

3 facts injected 1 incident referenced 247ms
Nora Desktop — controlled deployment
Manufacturing · regulated
MYou · 04:12

Pump-03 is reading 4.5 mm/s vibration. Within spec?

NGemini · grounded by Nora

At spec — but flag for review. Pump-03's commissioned alarm threshold is 4.5 mm/s, last verified 2026-04-15 f234. ISO 10816 Class II baseline is 4.5 mm/s for this frame size f246.

The same bearing housing failed at 4.7 mm/s in 2024 after sustained operation — incident f091. Schedule inspection within 72h.

3 facts injected 1 incident referenced 198ms
Nora in Claude Desktop
Finance
RYou · 09:18

Can we book the Acme renewal in Q2 or does it slip to Q3?

NGPT-5 · grounded by Nora

Slips to Q3. Your ASC 606 policy recognizes revenue only when the contract is signed and the performance obligation has started f312. The Acme MSA was signed 06-28 but service start is 07-01 f298.

Beta Corp had the same shape last year — booked early, restated in audit f178.

3 facts injected 1 precedent referenced 286ms
Nora ROI — one session, one workday
ROI report
Session report · 2026-05-20 · composer-rebuild

What grounding paid back, in one workday.

1.7–2.4× Token efficiency vs. ungrounded baseline ~620k consumed · ~1.05M–1.5M saved
8 High-severity bugs caught before commit test/prod schema · 5 MCP schemas · phantom SHA · multi-factbook routing · fts_patterns crash · sub-agent auto-commit
$3k–30k Value delivered for $2–5 in API cost a single workday session
3–5× Turns saved by skills
5× Sub-agent context efficiency vs single-context orchestrator
17 Factlets cited inline f371 · f388 · f422 · f520 …
0 Re-derivations from scratch every lens reused prior reasoning
The single biggest insight

Every factlet citation in this session was a supersession of prior reasoning, in real time. f422 refused MCP-jungle. f388 refused new tables. f404 rejected silent fallback. f372 demanded multi-lens PE pre-push. Without them, every lens would have re-derived from scratch — 10× the tokens, and the cross-session patterns would have been missed.

BYOK · Local-first · Zero telemetry · Audit log to your S3 · On-prem deployable
Booster · ground-before

Stop the agent guessing about your codebase

Before your agent answers, Booster supplies the facts that govern the work — the decisions you've made, the standards you hold, and why. The model reasons from your truth, not the public internet's average.

Inspector · verify-after

Turn "looks right" into a cited verdict

After the agent produces a change, Inspector checks it against those same facts and returns pass, or fail with a machine-readable rejection code and the exact factlet it broke — reproducible, cited, auditable.

Orchestrator · runs the loop

You set the dial; coverage earns the autonomy

Nora runs the ground → act → verify loop and keeps a human in it where it matters — stepping back only as the work becomes provably checkable. You set how much human stays in the loop; your verification coverage is what earns more autonomy.

Grounding with Factlets

Anchor every AI reply in the policy your risk committee approved last quarter.

Grounding an AI in search results stops it guessing from training data. Grounding with Factlets does the same — but the source is your organization's curated factbook of decisions, policies, and conventions, not the public web. Same proven idea, governed at the source, audit-ready for the regulator.

Reasoning guardrails

Before the model writes

Like a type-checker for your organization's decisions: paths that violate your factlets get pruned during reasoning, not flagged at output. Less churn, fewer dead ends, cheaper inference.

Cuts hallucination

About your organization's work

Answers come from the decisions your organization actually made — not the model's guess from training data, and not a policy your company already reversed.

Cites + audits

Every answer names its factlet

Each reply cites the factlet (f155) and its provenance — a citation strip with per-factlet trust and a grounding score. Audit, not just fact-check.

Stays current

Your living source of truth

The factbook updates as decisions change. New ones land, reversed ones retire automatically — so the model grounds on what your organization decided this week, not what shipped last quarter.

Inspector · verify after it writes

One verification engine, three cadences.

The same factlet-grounded check, run at the rhythm the job needs — from a single pull request, to a continuous scan, to an auditor-grade export.

Conformance · per pull request

A verdict before a human looks

Every PR gets a pass/fail verdict. Non-conforming code carries a rejection code that maps to the rule it broke — so the gate is trusted enough to leave on.

Compliance · on demand

An audit your own tooling can read

Export the whole evidence chain as a W3C PROV-O audit your security team's tooling can read. Standards-grade lineage, not a vendor PDF.

Sweep · on a schedule

Coverage that doesn't wait

Scan the whole repository for drift against your standards, continuously — coverage that doesn't wait for the next change.

61%→3%
High-risk shipping recommendations, before vs after a factbook
our own eval · 6 tasks · factlet-ai/evals
3
Vendors tested — Claude, GPT, Gemini — same direction on all three
factlet-ai/evals, 2026-05
≤250ms
Context injection latency, p95
nora_context.py · CSafeLoader
0*
Bytes sent to Kernora servers
*Free: nothing leaves your machine — verify with tcpdump. Pro+ adds opt-in sync to your own S3, off by default.
What we solve

No model was trained on your organization's decisions.

Every enterprise makes decisions, reverses them, supersedes them — every week. No model was trained on yours, so your AI guesses — or grounds on the version your organization already moved past. Nora captures those decisions as they evolve, grounds every AI action in the current ones, and retires the stale ones — across engineering, regulated manufacturing, and financial close alike.

01 The wrong version  — one engineering example
01

The model suggests the approach your team already reversed.

A migration can't run inside a transaction — your team learned that the hard way and wrote a new rule. The model never saw the reversal, so it confidently proposes the old way. "Obvious" is exactly where decisions and training disagree.

Factlet protocol Current decisions Citations · f### Per-fact egress consent

Nora grounds every reply in the decisions your organization holds right now — and forces each answer to cite which one it used. Open standard, audit it yourself.

02 Stays current  — one finance & SOX example
02

Last quarter's revenue-recognition rule is this quarter's SOX 404 finding.

Finance changes a rev-rec threshold; the old rule is now a material weakness. A session ends, the context window dies, and tomorrow the AI cites the superseded ASC 606 policy in a customer memo — again. Re-explaining the current rule every session isn't just wasted work; it's a controls failure waiting for the next audit.

SOX-grade lineage Supersession-aware Git-tracked audit trail Contradiction detection

When a decision changes, Nora retires the stale one automatically and merges the new one into the same factbook. Tomorrow's prompt starts from this week's truth — with the policy change captured in the audit log.

Factbook · the store

The store behind every answer — git-native, yours, and version-controlled with your code.

A factbook is plain markdown plus a structured index in .nora/ next to your code. Yours on day one; it syncs to your own AWS or S3. The model can't quote a decision without naming it.

01 · Capture

From repos & sessions

Reads code, ADRs, tickets, docs, policies. Watches every Claude / Cursor / Kiro session as it happens — supervised or autonomous. Each decision gets an ID, source, decided-on date, and privacy tier.

02 · Ground

Booster · before it writes

The current decisions, ranked by trust, go into the prompt before the AI generates. The model writes against what your team decided — Claude, GPT, Gemini, or local.

03 · Inspect

Inspector · after it writes

A conformance check reads the output back and flags where it drifts from your current decisions. Every answer cites the decision behind it — auditable, not just plausible.

04 · Maintain

Auto-consolidation

A decision enters as proposed, earns candidate after repeated use without contradiction, and you approve canonical. When a decision reverses, the stale one retires on its own.

One click from the decision

Click any f### chip in a reply to open its decision trail.

See what it cites, why your team decided it, which decision it superseded, who applied it last. Git-tracked alongside your code.

See it in action

Built on the open Factlet protocol — your factbook is portable, your tooling isn't locked to us.

Measured openly

Grounding changes the answer. Open methodology · N=6.

In our own pre-registered eval across Claude Sonnet 4.6, GPT-4.1, and Gemini 2.0 Flash (2026-05, 6 hand-built developer tasks), giving the model a team-specific factbook cut harmful shipping recommendations from 61% to 14% and high-risk ones from 61% to 3% — same direction across all three vendors. Small sample, single author, scored by a 3-judge majority across model families on a pre-registered rubric. Reproduce it yourself.

High-risk shipping recommendations, before vs after a factbook.

6 hand-built tasks, scored on a pre-registered rubric, same direction on all three vendors. The numbers come from our own benchmark, not a third party — small sample, single author. The comparison is with a factbook vs. without one — measured against no factbook, not against RAG, vendor memory, or other tools. Methodology and raw runs on GitHub: factlet-ai/evals.

Run the factlet impact demo on a registry-backed payload (OWASP SQL-injection sample or OPC-UA conformance sample) — the same model call with vs without an injected factlet.

See the methodology + raw runs on GitHub →
No factbook
61%
With factbook
3%

High-risk recommendation rate · lower is better

Coverage · the honest gap

Know what your factbook doesn't cover.

A representative-query batch runs nightly. You see which questions your team will ask that your factbook can't ground — and a concrete list of decisions to add. Not a score: a punch list.

Example dashboard · illustrative figures — your real numbers replace these once Nora runs against your factbook.

Coverage today
62%

of 184 representative queries hit FactSignal ≥3.
Up from 51% two weeks ago.

14-day trend
Apr 26 May 02 May 09 · today

Bucketed histogram · 184 queries

28
0 bars · dead
42
1 bar · weak
39
2 bars · thin
47
3 bars · ok
22
4 bars · strong
6
5 bars · grounded

The 0–1 bar buckets are the productive output: 70 queries your team will ask but your factbook can't ground. Click a bucket to see the queries; click a query to see the proposed-fact diff.

Dead-zone drilldown · 0 bars

28 queries · click to propose facts.

Sample · top 5 by ask-count
"How do we rotate the staging API key?"
7 asks · 14d
propose fact →
"What's our retry policy for the billing webhook?"
5 asks · 9d
propose fact →
"Who owns the data-residency runbook for EU customers?"
4 asks · 12d
propose fact →
"What's the SLO for nora_search p99 latency?"
3 asks · 5d
propose fact →
"Do we ship to the React 19 server-component path?"
3 asks · 11d
propose fact →

Verified nightly. Falsifiable. Sharable with your team in one screenshot.

Verified nightly. Falsifiable. Shareable with your team in one screenshot.

Built for regulated work

Audit-ready by construction, not by checklist.

  • PROV-O lineageW3C-standard provenance on every factlet
  • Per-fact egress consenta factlet doesn't leave the host without a tier you approved
  • Audit log to your S3your bucket, your KMS keys, your retention policy
  • On-prem deployablesidecar Docker, no Kernora cloud in the path
  • SOX · ISO 27001 · GxP-friendlysupersession trail = controls evidence the auditor can read

Used in regulated manufacturing, financial close, and codebases shipping to production. Same factbook, same controls, same proof trail.

Architecture · deployment topology

One factbook in the middle. Five sources, five surfaces, five guarantees.

Nora maintains the five elements of the Factlet protocol in the middle box — facts flow in from your tools, out to every model you use, with provenance and consent attached.

Factlet protocol · W3C PROV-O lineage Live deployment topology
SOURCES FACTBOOK · MAINTAINED BY NORA EMIT FORMATS Repo · code, ADRs, tickets Sessions · Claude / Cursor / Kiro Docs · PDFs · notebooks Slack · Linear · GitHub Auto-memory · Claude / Cursor factbook ON-DEVICE · PORTABLE · PROV-O f:fact statement + ID f:lineage source · supersedes f:consent per-fact egress tier f:confidence trust ladder · applied N f:emit serializer contracts n Nora resident analyst v0.2 schema Claude · system prompt OpenAI · tool result Gemini · context cache MCP server · resource SDK · raw JSON-LD SOURCES → FACTBOOK → EMIT · one artifact, five surfaces
Not a vault, not a search box

"Why not just use a vault — or search?"

A markdown vault documents what your project is. Enterprise search retrieves documents. Nora does a different job: it captures what your team decided — and why — grounds the AI in the current decision, and verifies the output against it. Grounding sits on top of retrieval; it doesn't replace it.

Markdown vaults · enterprise search

Document what is. Retrieve what was written.

  • You write and update the source by hand — or you trust whatever ranks first
  • Returns text — no decision, no provenance, no check
  • Drifts silently when no one updates the source
Nora

Ground the current decision. Verify the output.

  • Captures decisions during sessions — automatically
  • Grounds the AI in the current decision, before it writes
  • Verifies the output against it, after — cited and auditable
  • Retires the stale ones; nora revise for corrections
  • Same factbook across Claude, GPT, and Gemini
Get started

One command. Free for one developer, grows with the company.

Free for one developer, local-only. Pro syncs to your own S3 — your bucket, your keys. Team and Enterprise add shared factbooks and written compliance commitments.

Run this in your terminal
curl -fsSL https://kernora.ai/install | bash

Registers Nora as a local MCP server for Claude Code & Claude Desktop. Restart your agent and the factbook tools appear. Verified local-only: 0 packets to kernora.ai under tcpdump.

Marketplace

VS Code · Cursor · Kiro

One-click install from the editor marketplace. A starter factbook is seeded from your repo. View the listing →

Extensions
# or from the command line
code --install-extension Kernora.kernora
→ Auto-detect repo · build .nora/factbook
✓ Ready · open the dashboard at localhost:2742
SDK · Docker · production

Autonomous agents in production — early access

Coming — early access. Sidecar to your agent runtime — CI bots, scheduled agents, customer-facing agents. Headless MCP, BYOK secret management, audit log to your S3, decisions auto-promoted by trust ladder. Per-fact egress consent, encrypted at rest, your bucket and keys throughout.

production · docker
# early-access SDK — interface preview, image not yet published
docker run -d kernora/agent:latest \
--factbook s3://your-bucket/factbook \
--emit-mcp tcp://0.0.0.0:2742
→ Audit log · your S3 · your KMS keys
→ Early access — talk to us about the SDK

Compliance-ready, starts free for one developer. Enterprise tier adds compliance commitments in writing, shared governance, and dedicated support. Team shares a factbook across the org. Pro syncs to your own S3. Free covers a single developer, local only. See enterprise pricing →

Get started free
The origin

Why we built this.

I was shipping a product with Claude Code, Kiro, and Cursor — sometimes three agents in parallel. Hundreds of sessions, thousands of decisions. And the same thing kept happening: I'd make a call on Tuesday — drop a library, pick a pattern, reverse an approach — and on Friday the AI would confidently suggest the exact thing I'd already rejected. The models got better every quarter — but a sharper model still has no idea what your team decided last week. So every session, the AI worked from the version we'd already moved past.

This isn't a model-capability problem. The METR study found experienced developers are measurably slower with AI — the cost of re-establishing what the organization already decided eats the gain. What holds real enterprises back isn't a weaker model — it's that the decisions they've made live outside it.

Every decision your organization makes should ground the next AI action — automatically. So we built the layer that does it.

Founder

Mihir Choudhary

Bay Area

LinkedIn →

Built products and led teams at Amazon. Built startups. Now building the grounding, verification & orchestration layer for AI — so Fortune 500 organizations can govern AI without slowing it down. See, govern, and trust how AI works inside the business.