Trust & privacy

Own the context your AI uses.

Holoself keeps canonical context as local Markdown. It has no hosted account, does not require external AI or vector services, and does not silently update canonical self.

Read by default. Propose before writing. Preserve evidence and provenance.

Open-source preview
Safety guarantees

Designed for reviewable local control.

No required network calls

Core CLI behavior runs locally. No hosted account, publication, or external action is required.

No silent canonical writes

Projects receive read access. Accepted reusable knowledge goes through proposal, displayed diff, explicit confirmation, and validation.

No hidden source of truth

Markdown remains canonical. Indexes and packets are generated, inspectable, and safe to rebuild.

Privacy-aware resolution

Visibility, public-safety, sensitivity, excluded lenses, and supported field restrictions filter context.

Secret filtering

Credential-like filenames and common secret patterns are excluded from context and index output. Detection is defense-in-depth, not a vault.

Provenance retained

Context results identify sources. Approved claims record project, source file, evidence, confidence, visibility, and approval.

Threat boundaries

What Holoself does—and does not—protect.

RiskControlYour responsibility
Accidental canonical changeRead-only links, proposal states, confirmation, diffReview evidence and target before approval
Over-sharing across lensesVisibility and sensitivity filteringClassify content accurately; test public workflows
Secrets entering indexFilename and content pattern exclusionNever store credentials in context; rotate exposed secrets
Private files committed publiclyPrivate-by-default guidance and local pathsConfigure Git ignores and review staged files
Compromised local AI toolBounded context and project accessTrust tools granted filesystem access; Holoself is not a sandbox
Known limitations

Trust requires honest boundaries.

  • Holoself cannot prevent an authorized local tool from reading files outside its intended workflow.
  • Secret-pattern detection cannot identify every credential or sensitive passage.
  • Visibility metadata is user-authored policy; incorrect classification can produce incorrect access.
  • Semantic duplicate and contradiction analysis is heuristic, not proof.
  • There is no hosted sync, account recovery, encryption service, or remote access control.
  • The published npm package is a distribution channel, not a hosted Holoself service; context and operation remain local.
  • Optional sync or private Git storage inherits the security model of that service.
Verifiable proof

Inspect claims against code and evidence.

Privacy policy

Data boundaries, package contents, and local handling.

Open PRIVACY.md

Schemas

Machine-readable contracts for links, proposals, and context output.

Browse schemas

Tests

Behavioral evidence for links, lenses, privacy, proposals, indexing, validation, and safety.

Browse tests

Review trust before adding real context.