Skip to content
Arctite

Give AI the context it needs. Nothing it doesn’t.

Arctite observes the development sources you approve, assembles and redacts context locally, and lets you review the exact package before it reaches your AI provider.

Join the waitlistSee how it works
Arctite- aurora-project

workspace / aurora-project

Anthropic · ClaudeBYOK

Intent

Fix authentication race condition

inbox #142 · requested 2 min ago · feature/auth

Context readiness

86%

Observed

  • src/auth/session.ts
  • Git diff · feature/auth
  • Terminal · npm test → 2 failures
  • Browser context · 3 tabs

Sensitive data

2 items redacted locallyAPI keys · tokens
Connected

local engine · redaction on · windows 11

AI coding tools have a context problem.

Too little context

Models often miss what happened outside the current prompt - the branch, the failing test, the file that changed an hour ago.

Too much access

Autonomous tools increasingly request broad repository, terminal, and system permissions before they have anything useful to do.

Too little visibility

Developers usually cannot verify the exact payload that reaches the model - what was included, and what leaked with it.

Arctite puts a local context and control layer between your development environment and the model.

How Arctite works: observe, review, and act on approved context.

Observe

Approved sources gather relevant development context - locally, per project.

Review

Redact, freeze, and inspect the exact package before anything is sent.

Act

Turn approved context into implementation guides and gated agent cycles.

Observe only what you approve.

Every observation source is explicit, project-scoped, and shown on one screen. What is not listed here is never gathered - no keystrokes, no cookies, no screenshots, no mouse tracking, no browser storage.

Source policy is stored per project and travels with your workspace file.

OBSERVATION SOURCES

per project

Approved sources

  • Filesproject scope only
  • Gitdiff · log · branch
  • Terminal metadatacommands · exit codes
  • Browser contextactive tab · page title
  • Clipboardmanual share only
  • Notesproject notes

Never captured silently

  • Keystrokesnever captured
  • Cookiesnever captured
  • Screenshotsnever captured
  • Mouse inputnever captured
  • Browser storagenever captured

Permission set is explicit, project-scoped, and re-confirmed on change.

See the exact context before anything leaves your machine.

CONTEXT PACKAGE

FROZEN17 items4,810 tokens2 redactions

Files

├─ src/auth.ts · 1.1 KB├─ src/session.ts · 1.2 KB└─ + 15 more items

Git

├─ branch: feature/auth└─ diff: 41 additions · 17 deletions

Terminal

└─ npm test → 2 failures before fix

Browser

└─ 3 tabs · auth pages only
ProviderAnthropic
ModelClaude
Digest7e3b0f…a91c
Redactions2 secrets removed

This exact package, and nothing else, is what gets sent.

Frozen 23:14:08 · sha 7e3b0f…a91c

What you approve is what gets sent.

Turn approved context into an implementation path.

Each intent becomes a local guide built from the approved package - diagnosis, affected files, implementation, verification, and a known rollback path.

GUIDE · aurora-project

LOCALVALIDATED

INTENT #142 · 2 min ago

Fix authentication race condition

Diagnosis

Session lookup races with token refresh on startup. Two concurrent getSession() calls resolve against stale storage before the refresh settles.

Affected files

  • src/auth/session.ts
  • src/api/auth.ts

Implementation

src/auth/session.ts+1 · −1
1import { validatedSession } from "./validate"
2 
3export async function getSession() {
4- session = await getSession()
5+ const session = await validatedSession()
6 return session
7}

Wrap refresh in a single-flight promise so concurrent reads wait for the same settling token.

Verification

npm run verify -- auth
expect: 0 failures

Rollback

The package digest is frozen at approval. Restore returns to a known context package - no file writes happen without approval.

diff +14 −7 · verified locally

AGENT CYCLE

  1. Assemble context
  2. Review frozen package
  3. Write src/session.tsawaiting approval
  4. Run verification

Proposed write

+14 −7

src/session.ts

Type “approve” to execute this step

Agents should ask before they act.

Arctite validates generated plans locally and exposes execution one typed action at a time - never as an open-ended loop.

CONTEXT
PLAN
VALIDATE
PREVIEW
APPROVE
EXECUTE
VERIFY

Each transition is deterministic · each approval is bound to a frozen package digest

  • No arbitrary shell
  • No silent retries
  • No invisible file writes
  • No execution outside project permissions

No arbitrary shell.

No silent retries.

No invisible file writes.

No execution outside project permissions.

Every rule above is enforced locally and shown in the activity log - not promised in a policy document.

Arctite doesn’t ask you to trust it.It shows you.

Sources are gathered on your machine, assembled into a context engine, redacted and frozen, reviewed, and only then sent as the exact approved payload to a provider you chose.

YOUR MACHINE

all context processing happens here

Files
Git
Terminal
Browser
Clipboard
Notes

ARCTITE CONTEXT ENGINE

assembly · redaction · packaging

Redact
Freeze
Review
Approve

digest bound to approval

EXACT APPROVED PAYLOAD

4,810 tokens · digest 7e3b0f…a91c · 0 unapproved items

user-selected provider:OpenAIAnthropicOpenRouterLocal models

local models never leave your machine

Your repository isn’t our business.

Arctite is designed without an Arctite project-data cloud for developer context. There is nothing to upload to, and no hosted memory of your project activity.

Doesn’t need

  • Repository uploads
  • Hosted context history
  • Cloud memory of project activity
  • Keystroke capture
  • Cookies
  • Screenshots
  • Browser session storage
  • Silent monitoring

Built around

  • Explicit permissions
  • Local redaction
  • Local context storage
  • Exact payload previews
  • User-selected inference
  • Project-scoped access
  • Approval before writes

Context stays local until you explicitly approve a package for your selected inference provider.

Nothing leaves silently. Nothing is lost quietly.

Every approved package is frozen to a digest before it is sent. If a provider fails or a run goes sideways, the workspace returns to a known package - not a guess.

RECOVERY · SESSION SNAPSHOTS

LOCAL

Provider authentication failed.

Your context package was not sent. Nothing left the machine.

TimePackageTokensDigestStatus
23:14:08auth-race-fixpkg #1424,8107e3b0f…a91cverified
23:12:47auth-race-fixpkg #1394,69292c1f4…d4f7frozen
23:09:12auth-race-fixpkg #1315,032b0e4a3…3f2afrozen

Every approved package is frozen to a digest before it is sent. Restoring returns the workspace to a known context package - no file writes without approval.

Everything the system does is visible after the fact.

The activity ledger records every assembly, redaction, approval, and execution step - with its digest. What happened, when, and what it touched.

Append-only local ledger · readable by you, and only you

ACTIVITY · TODAY

local ledger · append-only

  1. 23:10:31

    Intent created

    “Fix authentication race condition” · feature/auth

  2. 23:11:02

    Context assembled

    6 sources observed · npm test → 2 failures

  3. 23:12:47

    Redaction applied

    2 items · API keys · tokens

  4. 23:13:15

    Review completed

    17 items confirmed · 0 unapproved

  5. 23:14:08

    Context package frozen

    17 items · 4,810 tokens · 7e3b0f…a91c

  6. 23:14:21

    Provider request approved

    4,810 tokens → Anthropic · Claude

  7. 23:14:39

    Plan received

    3 typed actions

  8. 23:15:04

    File write approved

    src/auth/session.ts · +14 −7

  9. 23:15:16

    Verification complete

    PASS · npm test → 0 failures

Experimental

Evolution Lab

Build and evaluate deterministic, task-specific graph modules locally inside a bounded interpreter. No filesystem, no network, no credentials - just candidates, scores, and a selected module.

Best module selected by score, then folded into a guide - never executed automatically.

EVOLUTION LAB · EVAL RUN 0x2F

EXPERIMENTALbounded interpreter

Flow

  1. Dataset
  2. Graph candidates
  3. Evaluation
  4. Selection
  5. Best module

Evaluation readout

local · bounded · deterministic

epoch 0 → 24score converging · best 0.94

generation

24

candidates

128

best fitness

0.94

evaluations

4,096

selected module: session-guard-v9 · fitness 0.94

sandbox:No filesystemNo networkNo credentialsNo process executionDeterministic evaluation

Questions, answered plainly.

If something here doesn't match what you need, the waitlist is the fastest way to tell us.

Is Arctite an AI model?

No. Arctite is a local-first control layer that sits between you and the models you already use. It assembles context from your project, freezes it into a verifiable package, and gates every generated change behind your approval. Arctite doesn't replace your model - it wraps it with context, control, and accountability.

Does Arctite send my code to a server?

No. Context is assembled, hashed, and stored on your machine. Nothing is uploaded for indexing, training, or storage, and every package you approve is bound to a digest that is verified before execution - locally.

Which models does Arctite work with?

Arctite works with the models and tools you already run - hosted or local endpoints alike. It doesn't dictate your stack; it layers deterministic context and approval on top of whatever you already use.

What is a frozen package digest?

Before anything executes, Arctite freezes the proposed change into a package with a cryptographic digest - file paths, diffs, and actions included. An approval is bound to that exact digest, never to a moving target, so what you approve is exactly what runs.

What happens if a package is rejected?

Nothing executes. The proposed change stays in its recovery snapshot, and every package in history can be restored at any time. Recovery snapshots preserve the state of your project before each cycle, so rollback is deterministic rather than guesswork.

What does “no arbitrary shell” mean?

Every execution is typed and permission-bound - a specific command with specific arguments, within project permissions. Arctite never hands a model a free-form shell in which anything could run.

Can Arctite work offline?

Core context assembly, hashing, and recovery run entirely on your machine. How much runs without a connection depends on the model and tools you point at Arctite - the layer itself doesn't upload your context anywhere.

Does Arctite replace my editor, terminal, or CI?

No. Arctite sits beside the tools you already use and leaves your git workflow, tests, and editor untouched. It layers context and approval on top of them instead of re-architecting them.

Is my context private even with a hosted model?

Your context stays on your machine, your approvals stay local, and redactions are applied before anything is shared. When a model call happens, only what's needed for that call is sent to the provider you chose.

How can I get access?

Arctite is preparing an invite-only private beta for Windows 11 developers. Joining records your email and CTA source; access will be offered in stages, and early users may be asked to share product feedback. No access date is promised.

Plans that stay out of the way.

Individual access is the private-beta focus. Team capabilities are planned and not yet available. Final pricing is not set.

Individual

PRIVATE BETA

For developers using Arctite locally.

  • Local context engine
  • Exact context preview
  • Guides
  • Agent Cycles
  • BYOK providers
  • Multiple projects
Join waitlist

Final pricing is not yet set

Teams

PLANNED

A planned edition for organizations requiring centralized controls.

  • Everything in Individual (planned)
  • Shared workspaces (planned)
  • Roles (planned)
  • Organization policy (planned)
  • Audit controls (planned)
  • Device management (planned)
Request access

Team access is not yet available

Windows-first private beta

Give AI context without giving up control.

Request invite-only access for Windows development workflows. Access will be offered in stages, and early users may be asked for product feedback.

We collect only this email address and the CTA source for private-beta access. See our Privacy Policy.