Vincent van Gogh, Wheat Field with Cypresses (1889), oil on canvas
security · data ownership

Your code and your memory never leave your infrastructure.

No proxy, no hosted service, no black box in the middle. The one thing that leaves: prompts still reach Anthropic’s models through the same Claude Code you already run, redacted first.

code stays in your repos, your infrastructure memory a git repo on your own disk secrets redacted before disk, before any model sync only to a remote you point it at

Mechanisms you can inspect, not certifications we ask you to trust.

scroll ↓

The three mechanisms

Three guarantees, each a real mechanism.

Nothing here rests on a policy promise. Each guarantee is something the software does: where the code runs, where the memory lives, and what gets stripped before it moves.

01runs in your stack

No proxy, no host.

Nothing sits between your machine and your work. Both organs are plain local tools (AgentHarness is a Claude Code plugin, the memory a CLI plus a capture hook), so neither introduces a server in the path.

Your agent runs where it always did, in the Claude Code your team logs into, against the models you already pay for. The harness reads each repo's AGENTS.md and drives the SDLC locally. Git operations use your GitHub auth. There is no middle tier to route code through, no account of ours that ever sees a diff.

installs into your Claude Code · your GitHub auth · your models · no relay, no telemetry

02the memory · local-first

Your second brain is a git repo you own.

Hyperception's memory lives in ~/.hyperception/vault/, its own local git repository on your disk. It is files you can read, diff, and delete.

Sync is opt-in and points wherever you say. Run hyp sync remote <url> once and the memory pushes to your private remote: a repo on your GitHub org, your GitLab, your own box. There is no Hyperception cloud it reports to. Move machines or hand a project to a teammate and the memory travels the way any git repo does, on infrastructure you control.

~/.hyperception/vault/ · plain files under git · hyp sync remote <your-url> · no vendor cloud

03redaction · before disk, before model

Secrets never reach the model.

Redaction runs twice: once before a captured session is written to disk, and again before any model reads it for digest. Both passes happen on your machine.

  • 1Before disk. The capture hook strips matches on every turn, so the raw transcript never lands in plaintext.
  • 2Before any model. The digest pass redacts again before the two-pass LLM ever sees the text, closing the gap for anything that slipped the first filter.

Patterns for tokens, keys, cards, and emails run first, and the pattern set is configurable: add your own for internal identifiers, hostnames, or customer data before you ever enable capture.

capture ▸ redact ▸ disk · digest ▸ redact ▸ model · configurable pattern set

In the interest of honesty

Mechanisms, not badges.

what this page does not claim

We make no compliance or certification claims. There is no SOC 2 report, no ISO number, no HIPAA or PCI attestation, no penetration-test seal, because none exist yet, and printing one that didn't would be the opposite of security.

And the honest exception to "never leaves your infrastructure": your prompts still reach Anthropic's models through the same Claude Code you already run. We add no new destination and nothing in between. That path is unchanged from how Claude Code already works, so the model you already use sees your prompts as it always has. (Redaction is a separate mechanism: it strips secrets from what Hyperception captures to memory, before disk and before the digest model, not from the prompts Claude Code sends Anthropic.) The promise is "no one new in the path," not "the model goes blind."

What we offer instead is architecture you can verify yourself: the code runs in your Claude Code, the memory is a git repo you own, and redaction happens on your machine before anything moves. Those are testable today. Need a DPA or a security review to clear procurement? Ask us. We'll work with your team, and formal attestations are on the roadmap, named here with real scope and date when they land, never implied.

Runs in your stack

Try it where it will always run.

A pilot installs both organs into your own Claude Code and runs them on one real repo for two weeks. Nothing leaves your infrastructure. You watch the mechanisms work on your own code.