Entity (identity)

Agentify discovers/indexes against Entity manifests. AgentDNA remains a separate canonical representation. Drafts stay in this browser only.

Website Agentifier

Proposal only. NOT authoritative. Trust stays unverified. No execution authority. Human review required before Apply.

Assess = Propose → Apply to editor → Phase-1 report (.md + .html) in one tap. Trust stays unverified. Report labeled SAMPLE. Fail-closed on empty paste. Saved to History (this device only, cap 25). Deterministic heuristics (mirror of engine/agentifier.py). Offline · no CDN · no paid APIs · PWA cacheable.

Discover (local semantic)

Tip: header Run guided demo loads SAMPLE registry, runs a sample query, then opens Execute with a preselected capability (no auto-authorize).

Discovery only. Discovery ≠ Authorization ≠ Execution. Trust ≠ authorization. Ranking is offline keyword scoring on this device. No network · no spend. Trust badges show stored status — never treat unverified as verified.

Primitive filter uses entity.extensions.semantic_primitives mappings from the Primitives tab. Classification ≠ verification. Unknown values are rejected (fail-closed). Shared relationships with another registry Entity matching the query can add a small ranking boost — never upgrades trust. Relationships ≠ authorization. Stale Entity state.freshness older than 7 days applies a slight ranking penalty — never invent freshness. State ≠ authorization.

Local registry (localStorage)

Save the current Entity into a device-local registry, import JSON, or load SAMPLE demo Entities (Calendly/Shopify-like pilots). Registry never leaves this browser.

Demo Entities are clearly marked SAMPLE. Discovery finds candidates — it does not authorize or execute. Use Execute on a capability (or the Execute tab) for dry-run → authz → simulate.

Execute (simulation)

SIMULATION only. Authorization ≠ Execution. Execution here is simulation only — never calls real external APIs (even Open-Meteo). Zero spend · offline · localStorage audit.

1 · Pick capability

2 · Inputs

This capability declares no inputs.

3 · Dry-run plan

4 · Policies reminder (read-only)

Policies ≠ execution permission by themselves. Reminder only — does not grant or deny authorization.

5 · Authorize (exact-plan, 10-min TTL)

Grant is bound to the plan fingerprint. Changing inputs/plan regenerates the fingerprint and invalidates any prior grant (fail-closed).

6 · Simulate commit

Local audit log

Append-only hash chain in this browser (localStorage). Stages: Plan → Authorization → Execution → Receipt → Replay. Mirrors engine/audit JSONL (prev + body SHA-256).

Import / verify portable bundle

Verification ≠ Trust ≠ Authorization ≠ Execution ≠ Ownership. Imported bundles are untrusted until Verify passes. A PASS only means the chain hashes/linkage are intact — it never grants trust, authorization, execution, or ownership.

Never collapses Discovery / Authorization / Execution. No network from this UI. Tampering any event fails Verify.

Capabilities

7 dimensions live in the manifest: identity, knowledge, capabilities, state, policies, trust, relationships.

Semantic primitives

Classification ≠ verification. Mapping a capability to Discover / Inspect / Prepare / Modify / Cancel does not verify trust, authorize, or execute. Native capability id and label are preserved — this is a non-destructive extension only (entity.extensions.semantic_primitives).

Mandate §8 primitives: Discover, Inspect, Prepare, Modify, Cancel. Assign one primary primitive per capability via the dropdown. Unknown values are rejected (fail-closed).

Suggest from kinds fills unassigned rows from capability kind (Discover/Inspect/Prepare/Modify/Cancel) — Classification ≠ verification. Save writes into the current Entity JSON under entity.extensions.semantic_primitives (map of native capability id → primitive). Does not erase capability id/label. Export downloads a portable mapping file. Offline · zero spend.

Relationships

Relationships ≠ authorization. Declaring a parent/child, owns, located_at, related_to, offers, or operated_by link does not authorize, verify trust, or execute. Links are a core Entity dimension (relationships[]) — not an extensions-only bag.

Suggested types: parent, child, owns, located_at, related_to, offers, operated_by — or enter a custom type. Target is an Entity id or name. URL and notes are optional.

Link map

Save writes the core relationships array (objects with at least type + target). Discover may apply a small ranking boost when a candidate shares a link with another registry Entity matching the query — it never upgrades trust. Offline · zero spend · mobile-friendly · no CDN.

State

State ≠ authorization. Availability, inventory, price, and freshness describe current Entity conditions — they do not authorize, verify trust, or execute. Persist as core state object. Freshness / updated_at are ISO timestamps when set — never invent freshness.

Key fields: availability, inventory, price, opening_status, capacity, health + custom key/value. Each entry can carry updated_at / freshness. Discover applies a slight ranking penalty when overall freshness is older than 7 days (configurable STALE_STATE_DAYS).

Save writes core state (not extensions-only). Overall freshness / updated_at update when you mark fresh. Offline · zero spend · mobile-friendly.

Policies

Policies ≠ execution permission by themselves. Declaring discovery, execution, cancellation, payment, privacy, eligibility, or geographic policy text does not grant authorization or execute. Persist as core policies object. Execute shows a read-only reminder before Authorize.

Suggested keys: discovery, execution, cancellation, payment, privacy, eligibility, geographic — plus custom label + value (mobile-friendly). Optional rules[] kept for fail-closed defaults.

Save writes core policies. Offline · zero spend · structured for mobile (label + text/value).

Trust

Trust ≠ authorization. Recording evidence or marking human-attested does not authorize or execute. Default remains unverified. Assess / Agentify / OpenAPI imports never auto-set status to verified.

Primary statuses: unverified · evidence_recorded · human_attested. verified is allowed only with evidence count > 0 and shows a strong warning — it is not a silent leap to production trust.

Evidence

Add/remove items: type (provenance | ownership | attestation | reputation | other), note, optional URL, recorded_at.

Save writes core trust (status + evidence[]). Discover prefers evidence_recorded / human_attested slightly over unverified — never hides unverified, never treats as authorized. Offline · zero spend · mobile-friendly.

Knowledge

Knowledge ≠ verified truth. Entries here are still evidence/review gated. Agentifier proposals are marked proposal and do not become verified truth by themselves.

Types: description · fact · documentation · schema · reference · other. Each entry: title/summary, body/text, optional URL, optional evidence note.

Save writes core knowledge as an array (legacy string / {summary,facts} normalized on load). Discover does a light keyword match on knowledge text. Offline · zero spend · mobile-friendly.

Live JSON + Validate

Report runs offline from the current Entity JSON (validated first). Labels SAMPLE or LIVE-INPUT — not business validation. Share pack downloads a store-method .zip (manifest + report .md/.html + MCP + A2A + OpenAPI sketch + README-SHARE.txt). Fail-closed if invalid. SAMPLE / not authorization / zero-trust assumption.

Projections are compatibility views only — not authorization or execution. Agent DNA light is a separate alongside projection (full AgentDNA deferred; Agentify ≠ AgentDNA). All four route through projectViaCanonical (mirror of engine/canonical_capability.py). MCP ≠ A2A ≠ OpenAPI ≠ OpenAI-tools: separate buttons, separate downloads (<id>.mcp.tools.json / .a2a.agent-card.json / .openapi.sketch.json / .openai.tools.json or .projections.canonical.json). Protected envelope includes canonical capability ids. Validate fail-closed first.

Local drafts (localStorage)

Evidence intake

Recording ≠ validation. Ingest / owner-verify / owner attestation never unlocks business validation (still 0/2 Hold). All entries stay proposed / unverified. No contact capture. No auto-validation. No Pass.

Inbound portal (download JSON, not email): docs/evidence/ · Schema: spec/owner_verified_evidence.v0.1.schema.json

Owner-verified evidence (founder paste)

Founder-submitted source · claim · artifact refs · timestamp · owner attestation. Labels: proposed / unverified. Fail-closed — never marks biz-val Pass.

Inbound evidence packet (reviewer JSON)

Evidence dossier = local projection of this browser ledger → always Hold (0/2). Never auto-Pass. Integrity-stamped · Verify ≠ Validation. Ledger chain verify (v0.3.33). Mirrors CLI agentify evidence dossier / dossier-verify / verify.

Local evidence ledger

Ledger stays in this browser’s localStorage (key agentify.studio.evidence.v0.1). Mirror of CLI python3 -m agentify evidence. Offline · zero spend · no CDN · always proposed / unverified.

Assessment History

Offline only. Entries stay in this browser’s localStorage. Never uploaded. Cap 25 to avoid storage blowups. Oldest drop off when full.

Select two entries → Compare (client-side only, no network).

After Assess or Generate report, an entry is saved (id, timestamp, entity, provenance URL if any, report .md+.html, SAMPLE flag, entity snapshot). Select two → Compare (side-by-side; stacked on phones) · optional download compare .md · Open snapshot · re-download · Share pack (.zip) · delete (confirm).

Five-stage interaction chain

Dry-run ≠ authorization ≠ execution ≠ validation. Discover → Understand → Capability → Authz → Execute. The wizard never auto-authorizes and never commits.

Offline · local only · business validation stays 0/2 Hold.