Give an assistant the background for a project, the examples for a piece of writing or the preferences for a trip. SEED brings the context you choose into the tools you use, with the original records kept in your vault.
Access begins through the SEED waitlist. This guide describes the intended launch experience. The verified server address, supported clients and exact tool schemas will accompany access; they are not published here yet.
What MCP adds
Model Context Protocol, or MCP, gives an AI application a standard way to discover and call external tools. A SEED connection lets a compatible assistant request useful context from an approved selection in your vault.
You choose the information and the receiving tool. The assistant can then use the available SEED capabilities as part of your conversation. It does not need the passwords or access tokens for the services from which you originally collected your records.
For a custom application that calls SEED directly, read the API guide. For the ChatGPT experience, see SEED in ChatGPT.
Start with a task
| Ask your assistant | Select in SEED | A useful result |
|---|---|---|
| “Catch me up on this project.” | Project notes, decisions and current constraints. | A concise brief with the reasons behind earlier choices. |
| “Find where I decided that.” | The relevant project or topic. | A dated source passage, with enough context to understand it. |
| “Use these examples when drafting.” | Writing samples you have chosen to share. | A draft informed by those examples. |
| “Plan a trip around what I like.” | Selected travel preferences and a budget range. | Suggestions that reflect your preferences without requiring your wider archive. |
| “How do I download my account data?” | No private collection is needed for public guidance. | The provider's export route, available formats and checks for the received files. |
Begin with a small selection whose contents you understand. A project collection can be useful before you have imported years of history. Read how collection and connections fit together, or use the download guides to collect records independently.
The connection journey
When access opens, use the SEED connection instructions for your supported client. The intended journey has five steps:
- Choose the receiving assistant. Start from its supported SEED integration or the connection flow in SEED.
- Sign in to SEED. Check the account and application shown in the authorization flow. Keep account credentials in that flow, outside the conversation.
- Choose the material. Select a reviewed brief or a defined collection. Check the allowed operations and the duration.
- Approve the connection. Review what the recipient can receive before access is granted.
- Try a specific question. Ask something the selected records can answer, then inspect the returned sources.
An assistant asking for more information does not enlarge its permission. Additional records, operations or recipients need a new sharing choice.
Choose one task or an ongoing collection
A brief for one task is a fixed selection, such as the background for a meeting or the material for a chapter. Review its contents before sharing it. An expiry limits future retrieval of that package.
An ongoing collection supports recurring work. A project assistant, for example, can use the collection as its notes and decisions evolve. The authorization experience needs to make clear whether new or updated items become available under that connection.
In both cases, distinguish the permission you granted from the records actually returned. A connection gives an assistant a defined opportunity to retrieve context; it does not mean the entire selection has already been sent.
What an assistant can request
The launch design centers on a small set of retrieval capabilities. These are descriptions of the intended jobs, not the names of currently published tools.
| Capability | What it is for |
|---|---|
| Connection information | Understand the active account, authorized selection and available operations. |
| Search approved context | Find relevant material within that selection. |
| Read a source or excerpt | Inspect the evidence for a result, subject to the same permission. |
| Prepare context for a task | Return a concise brief from permitted material with its supporting sources. |
| Collection guidance | Explain a provider's export route and what to check after downloading. |
Clients should discover the tools the connected server actually offers. Do not hard-code names from an illustration or assume every connection has identical capabilities.
Reading context is distinct from writing to the vault. A correction or a new memory should remain a reviewable proposal until the appropriate user action confirms it. Access to a story also does not authorize an agent to publish it, make commitments in your name or generate your voice.
Read the evidence as well as the answer
Useful personal context needs more than a sentence. The intended response model keeps these details together:
- The information: a fact, preference, decision or relevant passage.
- Its source: the record or records supporting it.
- Its dates: when the event happened, where known, and when the record was captured or updated.
- Its standing: whether it was stated by you, found in a source, inferred or confirmed after review.
- Its coverage: what part of the approved material was searched and whether the result is incomplete.
“The review is on Friday” and “a note from last month proposed a Friday review” mean different things. Ask for the source when the distinction affects your decision. A missing result can mean the selected records do not contain the answer, an import is still incomplete or the wording needs a more precise search.
A source reference does not grant access by itself. Opening or fetching it must still respect the current connection and the reader's permission.
Keep context current
Correct a mistaken fact in SEED so future requests can use the reviewed version. Retain the historical source where it explains what changed. When a preference changes, the old record may still be useful history even though it should no longer guide a current recommendation.
An answer already delivered to another assistant will not rewrite itself. After a correction, request fresh context and check the date and source version rather than relying on an earlier chat message or a cached brief.
The same principle applies when a source is removed from a shared collection: a summary built from that source needs to be reassessed before it is disclosed again. Sharing a summary should not become a way to share excluded records indirectly.
Resolve a connection problem
The exact client messages will depend on the supported integration. Use the underlying situation to choose the next step.
| Situation | What to do |
|---|---|
| SEED is not offered in your client. | Check the supported-client instructions and any workspace administrator requirements provided with access. |
| Sign-in expired or the account is wrong. | Reopen the supported authorization flow and confirm the intended account. |
| The assistant cannot access a record. | Check the approved selection. Add information only if you want that recipient to receive it. |
| Search returns no useful answer. | Check the collection, import status and dates; try a more specific question. |
| The answer uses an old fact. | Review the current record in SEED, request fresh context and inspect its sources. |
| A source was deleted or is no longer shared. | Let the answer show that its evidence is unavailable; do not keep treating the old citation as current proof. |
| A request is delayed or limited. | Follow the integration's retry guidance. Avoid repeated automatic calls while a request is already running. |
| The client supports tools but not an optional interface. | Use the documented tool flow or a supported SEED page for the corresponding choice. |
For support, share the client name, the time, the operation and any request identifier offered by the integration. Keep access tokens and private source passages out of routine support messages.
Compatibility means a tested journey
MCP support is one part of compatibility. The client also needs the transport, authorization and tool features used by the SEED connection. Organization settings or the client's account plan can affect access.
The supported-client documentation supplied with launch should identify the connection method, available operations, account requirements and any optional interfaces. A working custom connection and an approved public directory listing are separate distribution routes.
For your first check, use a harmless project note. Retrieve it, correct it, request the updated version and end access. That demonstrates more than seeing a server marked “connected.”
End access when the task is finished
Revoking a connection stops future retrieval through that permission. It does not erase copies or answers already received by the external service. Disconnecting a source account, revoking an assistant's access and deleting records from the vault are separate actions.
SEED operates the core intelligence that organizes the vault on its own infrastructure. Connecting an external assistant is a separate disclosure choice: that service processes the approved response under its own terms. Some data or uses may also be restricted by the source or the receiving platform.
Review permissions and privacy · Explore the integrations · Join the waitlist