Surface coverage matrix

Gigabrain exposes its memory engine through four surfaces:

• Local MCP — the gigabrain_* tools registered for Codex / Claude / Hermes (lib/core/codex-mcp.js).
• Remote MCP — the read-only-by-default Streamable HTTP profile for Claude and ChatGPT (lib/core/remote-mcp.js).
• CLI — the gigabrainctl verbs (scripts/gigabrainctl.js).
• HTTP — the local /gb/* JSON API used by the web console and the Spark↔Gigabrain bridge (lib/core/http-routes.js).

The matrix below maps each capability to the surface(s) that expose it. It is a deliberate, asymmetric design: write/agent-loop capabilities live on MCP, operator maintenance lives on the CLI, and read/inspection surfaces are mirrored to HTTP for the console. A blank cell is an intentional non-goal, not an oversight — the rationale is in the notes.

Legend: ✅ first-class · 🟡 partial / indirect · ⬜ not exposed (by design or planned).

Capability	Local MCP	Remote MCP	CLI	Local HTTP	Notes
capture	✅ gigabrain_remember	⬜	🟡 sync-hosts, import-openclaw	🟡 POST /gb/suggestions	Broad remember-intent writes stay local. Remote MCP never exposes gigabrain_remember.
recall	✅ gigabrain_recall	✅ gigabrain_recall	✅ orchestrator explain	✅ POST /gb/recall, POST /gb/recall/explain	Remote authorization runs before ranking, exact server/token scopes are intersected, paths are redacted, and a receipt is recorded.
checkpoint write	✅ gigabrain_checkpoint	🟡 opt-in gigabrain:checkpoint	⬜	⬜	New checkpoints are immutable episodes. Durable candidates create proposals, not memories.
checkpoint list/get	✅	✅	⬜	⬜	Exact pagination and direct reads repeat scope checks; semantic recall is not used as enumeration.
claim propose/review/decide	✅	🟡 review by default; propose/decide require narrow scopes	⬜	⬜	Terminal decisions are append-only. Authenticated authority cannot be supplied by a tool argument.
receipt write/get	✅	🟡 get by default; write requires gigabrain:receipt	⬜	⬜	Recall receipts hash queries and remote responses omit local paths.
adjudications	✅ gigabrain_adjudications	⬜	✅ review adjudications	✅ GET /gb/adjudications	Existing arbitration inspection is not in the initial remote allowlist.
beliefs-as-of	✅ gigabrain_beliefs_as_of	⬜	✅ review beliefs-as-of --at <iso>	✅ GET /gb/beliefs-as-of?at=<iso>	Existing bitemporal inspection remains local in the first remote profile.
review-queue	✅ gigabrain_review_queue	⬜	✅ review queue	✅ GET /gb/review-queue	Existing global queue is not exposed remotely; typed claim review is the remote review surface.
watch	⬜	⬜	✅ watch (watch --install-hook)	⬜	Filesystem lifecycle is operator-only.
handoff	🟡 gigabrain_export_brief	⬜	✅ handoff	⬜	Full reports and exports stay outside the remote connector.
legacy checkpoint migration	⬜	⬜	✅ migrate legacy-checkpoints	⬜	Deterministic, idempotent, zero-promotion backfill; dry-run creates no tables.
legacy table drop	⬜	⬜	✅ migrate legacy-drop	⬜	Guarded operator lifecycle action.
vault	🟡 via recall	🟡 via authorized recall	✅ vault sync, vault status	🟡 via recall	Vault chunks remain read-only evidence and never become beliefs.

Read across, write narrowly

Two patterns explain the shape of the matrix:

1. Local inspection capabilities are mirrored. adjudications, beliefs-as-of, and review-queue are read everywhere — agents inspect them via MCP, operators via the CLI, and the console via HTTP. These are the trust-control-plane surfaces and benefit from being uniformly reachable.

2. Writes and lifecycle stay on their owner surface. Broad memory capture stays local; remote writes require a feature flag plus a narrow OAuth scope; watch and migrations are CLI-only; raw capture is an MCP write while the CLI only imports and HTTP only ingests structured suggestions. This keeps the blast radius of each mutating path aligned with the actor that owns it.

When you add a new capability, record its row here in the same PR so the asymmetry stays a deliberate, documented choice rather than drift.