Thegram Quickstart: Connect Cloud Memory, Then Read Safely · r11

Archived revision r11, recorded 2026-10-16T01:26:29.960Z by portal-product-quickstarts. Current page: Thegram Quickstart: Connect Cloud Memory, Then Read Safely.

Thegram quickstart

Choose one of the three Cloud connection paths below. A running Mac app, desktop socket and DMG are not prerequisites.

Choose the connection

Client	Authentication	Setup
ChatGPT chats and dots	OAuth	Install the published Thegram plugin and complete the provider account flow.
Claude	OAuth custom connector	Add the Cloud MCP URL in Claude Settings → Connectors. This is not an official directory listing.
Claude Code, Codex, OpenCode, OMP	Scoped API key	Install the shared coding plugin, issue a key in Thegram Settings, then connect privately.

Coding agents require Node.js 22+ and coding plugin 3.0.0 or later. Follow coding agent API setup. For ChatGPT and Claude, follow OAuth setup.

Storage and access

All current agent connections use Thegram Cloud. Local-only files are not uploaded or exposed automatically. API keys authorize selected vaults and separately selected Drive accounts, intersected with current membership and organization policy. A vault grant is not a document-only sandbox.

The independent API-key authority preserves credentials separately from Vault images, using its own sealed storage and lifecycle. Document requests still use https://vault.thegram.one/mcp; Vault restarts can briefly interrupt requests without erasing keys.

One type, one dashboard

Core 2.0.1 resolves one dashboard per declared document type. The client renders its single ResolvedView/1 result; it does not choose another ViewSpec or compute its own recommendation.

Document type	Dashboard
schedule@1	Journal
project@1	Project Tracker
knowledge@1	Knowledge

Without a template declaration, Core recommends a dashboard and the dashboard is read-only. If no recommendation is eligible, it falls back to Project Tracker, still read-only. An unknown id/version or malformed declaration has no dashboard: show the document and Core diagnostics without automatically choosing another type.

A folder opens one home screen. There are no per-document screen tabs, sidebar screen lists, ViewSpec switchers or search/filter/sort/group query toolbar. Document and dashboard are representations of the same file, not separate copies.

Templates still choose a type when creating a document, validate governed writes, supply AI authoring rules and define typed actions. A read-only recommended dashboard does not forbid ordinary document editing through the normal authorized write path.

For declaration and writing rules, see View templates.

First check: read-only

Discover the live tool contract with gram_capabilities and gram_projects. Use an exact returned mount key as folder for gram_brief, then read an existing path from its docs list. Do not guess tool fields from a historical reference capture.

gram_capabilities({})
gram_projects({})
gram_brief({ "folder": "<project folder or mount key returned by gram_projects>" })
gram_read({ "path": "<exact document path in gram_brief docs>" })
gram_read({ "path": "<same document path>", "section": "<heading returned by gram_read outline>" })

Expected: the permitted mount and existing documents are listed, and gram_read returns an outline or selected section. No write is needed to check connectivity. If the vault is empty, stop at the empty listing rather than creating data as a test.

With a write-scoped coding key, the first normal agent session may initialize that repository’s empty typed memory document. Worktrees share it. Read-only keys never create it; the repository brief is context selection, not an additional server-side document permission boundary.

Optional first save and search

Keep installation and the checks above read-only. Save only after the person explicitly asks for this save, grants write access, approves the exact new path, and confirms that it does not already exist in that authorized folder. Never create a mount, overwrite a document, or use a write as a connection test.

Use the actual gram_create input shape; base_revision null creates a new document. Replace the path with a new path under the exact authorized project folder or mount, and use only text the person approved:

gram_create({
  "path": "<authorized-folder>/setup-notes.gram",
  "title": "Setup notes",
  "base_revision": null,
  "content": "# Setup notes"
})

Expected: the tool confirms the new document and returns its path/revision. Read it directly; search the same authorized folder with gram_recall when indexing is available:

gram_read({ "path": "<exact path returned by gram_create>" })
gram_recall({
  "query": "setup notes",
  "folder": "<same authorized folder or mount key>",
  "top_k": 5
})

Expected: gram_read returns the saved document; gram_recall returns matching indexed material or no result until indexing catches up. Both retrieval calls are read-only.

Sources: https://thegram.one/install/agents.md · https://thegram.one/install/chatgpt.md · https://thegram.one/agent-memory/

Source and renderer evidence

Canonical source SHA-256

5ed2a602b802ea2e281917618a193d5d49d132a2d1dc391abc68d58857281144

Selected revision

r11 · sha256:9281a4dbd4a9a1d23ffe55cd9a88f9a8b14b8d2ab59ea06cc099d83ac7a0e7af

Local renderer

configured local Thegram renderer · API thegram-web-render/1

Core source

git+https://github.com/clab-one/thegram-format.git?rev=b16489e1d0d77d38f82d2f61776524737aa31557#b16489e1d0d77d38f82d2f61776524737aa31557

Canonical source: 11.gram