{ For Claude Code }

Your Claude, with a memory and a conscience.

Nogra gives Claude Code a bounded memory that learns you — and a verify-before-done discipline that shows receipts. Installed inside the real Claude. One line. Nothing leaves your machine — until you turn on sync.

Your model is already paid for. We just sync your brain.

Open sourceApache-2.0PyPI + npmyour Claude plan
claude code — your project

nogra ✓ verifiedbrief-2026-07-07 · receipt saved

{ Verification }

Evidence or it didn’t happen.

That verification block everyone’s hand-rolling into CLAUDE.md? It’s a product. It’s been shipping since June. One install — approve, run, verify.

The field’s own playbooks bake it in now: the worker isn’t the grader.

# CLAUDE.md

- ALWAYS verify before saying done

- only report work you can show evidence for

- no really, ALWAYS…

↑ hand-rolled and hoped for

✓ shipped as a product — approve · run · verify

{ Memory }

Hoarding is not memory.

Your agent doesn’t need to remember everything. It needs to remember what matters. Nogra’s memory is two markdown files you own, loaded into every session — bounded at 2,200 characters that MUST consolidate. Sharper agent, smaller bills.

MEMORY.md — 2,200 chars · USER.md — 1,375

consolidating

+ kept: what actually matters about you

- dropped: 40k tokens of transcript noise

loaded into every session — deterministically

✓ a theory of you — not a context bill

{ Autonomy }

Who checks the work?

Your agents now push code while you sleep. Nogra is the gate between “the agent says done” and “it ships” — brief, GO, evidence, verify. Human in the loop, receipts in the ledger.

03:12 — agent finished 3 tasks while you slept

gate: brief · GO · evidence

checked against the plan you approved

✓ verified — receipt in the ledger

{ How it works }

The model proposes. You decide — on evidence.

01

Brief

The work gets a short plan first — scope, stop criteria, what counts as evidence. Locked before anything runs.

The prompt asks. The brief binds.

02yours alone

GO

You read the brief. You say GO. Nothing dispatches without it — a brief is not a GO.

A goal is not a green light.

03

Evidence

The work runs inside the approved scope and comes back with receipts — files, tests, results. Not vibes.

Missing evidence is a result.

04

Verify

An independent check against the brief you approved — not the agent grading its own homework.

The model that wrote it does not sign off on it.

{ The first five minutes }

From install to first receipt.

No API key, no Python, no config file. Claude Code is the only prerequisite — the rest is five minutes.

min 1
claude plugin marketplace add nograai/nogra-claude-marketplaceclaude plugin install nogra@nogra-claude
min 2
/nogra:setup

the thin hub appears: .nogra/ + CLAUDE.md + empty inbox/ and projects/. Nothing more.

min 3
/nogra:brief “your first real task”

scope in, scope out, stop criteria, what counts as evidence. A brief is not GO.

min 4
GO

a scoped executor runs inside the approved boundary — and nowhere else.

min 5
evidence returns

verified against the brief, not the agent’s word. Receipt in your ledger.

✓ ran end-to-end on a stranger's Mac this morning — zero prerequisites, receipts in the ledger.

0

credentials touched. Auth stays first-party — it never sees your login.

2,200

characters of memory — bounded, and it MUST consolidate.

$20

is $20 for everyone, forever. No lifetime deals, no discounts.

{ The missing row }

Every agent ships the same table.

One row never makes it in.

any agent harness
nogra
Install
one-liner, then onboarding
two commands, inside Claude Code
Persistent memory
remembers everything, no receipts
bounded + receipts you own
Model access
your keys, credits on top
the Claude plan you already pay
Your credentials
handled by the harness
never touched — auth stays first-party
Autonomy
approval prompts, no plan to check against
gated — brief, GO, evidence
Who checks the work?
✓ independent verify · receipts in the ledger

Caught green-handed — our own ledger

[gate]classifier drifted mid-run — caught before ship, not after.

[gate]11 raw permission asks in one morning → one deterministic rule.

[verify]a stale build would’ve shipped the bug to strangers — caught pre-publish.

[gate]an agent tried to arm its own gate — escalated to the human, deterministically.

Everyone’s agents drift. Ours too.
The difference: ours leave receipts.

{ Pricing }

Your model is already paid for.
We just sync your brain.

No credit treadmill. No middleman between you and your model. Your Anthropic plan stays yours — Nogra carries zero inference cost.

Free

$0

local, forever

  • The whole workflow — brief, GO, evidence, verify
  • Bounded memory that learns you
  • Local-first: markdown files you own
  • Runs on the Claude plan you already have

Free is a product decision, not a disclaimer.

Install free

Sync

continuity

$20/mo

everything in Free, plus your brain everywhere

  • Your memory synced — desk, phone, always-on daemon
  • Sync without surrender: your key, your data
  • ✓ verified receipts, wherever you work

$20 is $20 for everyone, forever.

Join the Sync waitlist

Team

$49/mo

one brain, shared

  • Everything in Sync
  • Shared memory across your team
  • Shared briefs, shared ledger, shared receipts
  • One source of “actually done” for everyone

No tiers within the tier.

Join the Team waitlist

Enterprise — managed agents, self-hosted, on your terms.

Talk to us

{ FAQ }

The straight answers.

Cost, credentials, data, and the exit — the questions people actually ask, answered without spin.

No. Nogra is a Claude Code plugin — the same agent autonomy you'd pay API rates for, running on the subscription you already have. No second meter, no borrowed login, nothing to forward. Your Claude, doing more.

The model proposes.
Nogra proves.

Don’t trust. Verify — independently.