ChatGPT and Claude web

Connect either hosted chat to Grove's remote MCP server. No CLI, local plugin, browser extension or API key is needed. This is the custom-connection/private-beta path, not a public-directory listing. Your chat plan and workspace policy must allow custom MCP connections; an organization administrator may need to add it.

https://api.usegrove.cloud/mcp

For self-hosted Grove, use your own publicly reachable HTTPS API URL plus /mcp, with its matching website and OAuth issuer. Do not put tokens in that URL. See operator prerequisites.

ChatGPT

1. Open Plugins, then Add → Add custom MCP server. If the option is unavailable, check current developer-mode guidance and your workspace's permissions; do not purchase/change a plan just to follow these instructions without checking eligibility.
2. Name it Grove, select Server URL, and paste the endpoint above.
3. Select OAuth. Under advanced OAuth settings, use Dynamic Client Registration (DCR), with the discovered openid offline_access scopes. Grove's current issuer does not advertise CIMD; do not select it. Review the risk warning, create the custom plugin, and sign in to your Grove account.
4. Approve its Grove access as described below. Then return to ChatGPT, refresh the connection if necessary, and select Grove in a new conversation (for example by typing @). Interface labels can change; OpenAI's current guide is the reference for host-side controls.

Claude web

1. Open Customize → Connectors, then Add connector → Add custom connector. Choose Web if asked. Team and Enterprise administrators may need to add the connector for their organization before members connect their own accounts.
2. Name it Grove, paste the endpoint, and continue.
3. Choose Sign in now and Register automatically (DCR). No manually copied client secret is needed. Add the connector and sign in to Grove.
4. Approve its Grove access below, return to Claude, retry if necessary, and enable Grove in the conversation's connectors menu. See Claude's current setup guide if your dialog has a different layout. This is separate from the Claude Code plugin.

Approve the right connection

OAuth establishes identity, not permission to read your notes. You can scan the tool catalog while approval is pending, but note tool calls remain denied.

1. Open Grove → Settings → Connections, using the same Grove account used in the host's sign-in. Click Refresh connections if the new entry is missing.
2. Find the pending OAuth client by its client ID and creation time. A tool's pending-approval error links the exact connection and includes its authenticated client ID. A user-chosen name is not proof that a connection belongs to a provider. If several entries are ambiguous, do not guess or approve them all.
3. Select Approve, name the connection, and review its permissions. New connections start with every permission and scope checkbox selected: Read notes, Save and update notes, Delete notes, Personal and All current and future notebooks. Uncheck any access you do not want before approving. Read-only and selected-notebook access still work. Existing grants retain their saved choices; notebook nesting does not inherit permissions.
4. Approve, return to the host, and retry. Website approval alone does not prove the host is working: complete the synthetic save/recall check below.

Approval/edit forms carry a state fingerprint checked while the connection is locked. A newer scope edit or revocation rejects a stale save without restoring old permissions. Close the dialog, refresh and review again; normal tool discovery or last-contact updates do not invalidate the review.

To stop setup, cancel the host's consent flow. If an unapproved connection was already created, revoke that exact entry in Grove; it has no note capabilities. No approval link or client ID alone authorizes a grant. Changing accounts requires signing in again with the intended account, not approving another account's entry.

Check save and recall

Use a harmless example, not credentials or private information:

Use Grove to remember this preference in my personal notes: "I prefer concise answers with one worked example."

Check the reported saved ID and indexing status, then start a new conversation with Grove enabled:

Use Grove to recall my personal preference about answers and examples.

Try the same recall in the other connected host or an existing coding-agent integration. All use the same account's notes, subject to each connection's scope. Use update_memory for partial corrections; preserve the exact ID/version and reuse the operation ID on retries. Generic MCP saves keep the existing api source label unless the caller provides another supported provenance value; labels are not authenticated host identities.

Permissions and limitations

• Server instructions and argument docs share the note-writing policy: concise, focused Markdown and information-preserving complete-body edits. The host still chooses when to expose/call tools; this is not guaranteed automatic capture or a server-side readability validator. Refresh cached metadata after the matching API is deployed; no new connection/grant is needed for wording alone.
• The same 14-tool contract powers remote MCP and packaged integrations. Tools still enforce grants on every request. Tool annotations and host confirmation dialogs are not permission checks.
• Hosted OAuth cannot administer account credentials, public sharing or notebook subtrees, or read Grove's saved chat history. A listed owner-only tool still refuses the call. Notebook creation needs explicitly granted all-notebook write access. Uncheck all-notebook access if you do not want to allow it.
• Agent-initiated saves must use capture: automatic; requested saves use explicit. Grove's server preferences still apply. A local coding-agent pause or environment variable is not a hosted-chat privacy control.
• The host chooses when to call tools. This does not install session-start hooks, read every conversation, capture raw transcripts or guarantee automatic recall.
• Notes returned to a host have left Grove. That host's retention/privacy policies apply, and revocation cannot retract already delivered text. Grove is not end-to-end encrypted. Deletion hides retained notes/history; it is not erasure.
• No custom embedded UI, deep-research adapter or public-directory submission is included. Disconnecting a host does not delete your Grove notes; revoke access in Grove to stop future requests.

Troubleshooting

Never paste access tokens, client secrets or private notes into support reports. Use the support page with a sanitized error, host/date and request ID if available.

Verification and release status

Source work reuses existing OAuth/grants and requires no migration or new secret. Deploy the matching admitted API before the website instructions through the normal release path; merging source does not deploy it. Public listings remain separate.

The discovery rows above describe the initial pre-authorization checks. Later production 1.11.3 tests completed actual hosted OAuth and synthetic note operations: both hosts denied note content before separate Grove approval; ChatGPT saved, recalled in a fresh chat, inspected, corrected and soft-deleted one synthetic note; Claude read that same correction and exercised its own synthetic save/update/history and owner-only denial. ChatGPT also verified read-only, notebook-scope and revoked-connection denials. These bounded results do not claim an exhaustive refresh/expiry matrix or directory approval. Reviewer evidence and recording.

Local browser evidence and actual gate results are recorded in delivery and the reviewed gallery.

Official directory submissions

OpenAI's 1.11.3 package is submitted and In review, not approved or published. Claude's tested draft is held for written clarification of current Software Directory Policy §1.G for explicitly requested external preference notes and cross-session recall. Do not claim a policy exception or submit a false compliance acknowledgement. Until a listing is published, use the custom-connection paths above. Review each directory's current eligibility and submission requirements. Prepare descriptions/icons, privacy and support URLs, representative prompts, a synthetic reviewer account and repeatable OAuth/permissions tests. Evaluate stable client registration for scale rather than assuming the issuer supports CIMD. Submission, publication, provider changes, embedded UI and deep-research adapters remain distinct decisions. Operators can serve OpenAI's public domain proof with the optional API configuration; this does not authorize note access or satisfy the real-host/reviewer test gates.

The metadata follow-up publishes each existing tool's human-readable title both at the top level and in annotations.title, and keeps each description focused on its own operation rather than recommending another tool. Names, schemas, execution, safety annotations and owner-only authorization do not change. Deploy the admitted source before reconnecting/rescanning the existing directory drafts; a metadata fix is not policy clearance or permission to create duplicate listings.