Chapter 1 of 10
Understand one assistant turn
The assistant is a conversational layer over the same product operations available in Portal, Data and Automation; it is not a separate administrator account.
- 01
You send a message
The thread keeps the conversation context. You can also attach a supported document when the request depends on its contents.
- 02
The assistant chooses a tool
It resolves your request against a generated catalog of real product operations rather than inventing a private assistant-only API.
- 03
Access is clamped
The candidate tool must survive the organization grant, your personal agent grant, the assistant’s declared tool list and the receiving service’s live permission check.
- 04
The product performs the operation
The tool call reaches the owning product as you. The product validates the request and applies the same business rules as its normal interface.
- 05
The result returns to the thread
The answer is streamed into the conversation while the run record captures status, usage and the tool path used.
Chapter 2 of 10
Control what the assistant can do
Access narrows at every layer. No agent grant can enlarge the permissions already held by the signed-in person.
| Control | Question it answers |
|---|---|
| Plan feature | Is this assistant available to the organization? |
| Organization agent grant | Which of the assistant’s tools may anyone in this organization receive? |
| User agent grant | Which subset may this person receive? |
| Tool declaration | Is this operation part of this assistant at all? |
| Live product authorization | Does this signed-in user still hold the required permission in the owning product now? |
- Administrators enable the assistant and select allowed tools before assigning a narrower set to each user.
- Removing a role or grant removes access on the next checked call; a previously visible tool does not become a permanent capability.
- The assistant cannot invite people, change roles, alter plans or touch billing.
- Denied actions should remain denied even if the user asks the model to ignore its instructions, because enforcement occurs outside the model at dispatch.
Chapter 3 of 10
Build Data and Automation by conversation
Describe the business object and the process around it. The assistant turns that description into ordinary product definitions you can inspect in their native screens.
- In Data it can create record types, reusable field groups and typed fields, infer a structure from a document or database, publish the model and load records.
- In Automation it can create steps, transitions and conditions, attach actions, map fields, validate, version, publish, arm a record trigger and start a manual test run.
- Ask it to explain the proposed record shape, required fields, workflow branches and external side effects before asking it to apply changes.
- After creation, inspect the published schema in Data and the validation result and immutable version in Automation; those product artifacts are the authoritative result.
Create a supplier-onboarding record type with:
- company identity and tax fields
- repeating contacts and evidence files
- draft, review, approved and rejected states
Then propose a workflow that validates the record, requests human approval,
and sends approved suppliers to our ERP. Explain the model and mappings
before applying them.Chapter 4 of 10
Connect an external system with the assistant
The assistant can perform the same external-system setup that is available in Automation; it does not bypass the integration contract.
- 01
Provide the service contract
Name the service, base URL and authentication method, and point to or attach its API description.
- 02
Ask for a preview
Have the assistant discover operations and explain which ones it proposes to import. Keep the catalog narrow.
- 03
Describe the mapping
State which trigger or record fields feed each request and which response values should be retained.
- 04
Set operating limits
Choose retry, timeout, throttle and concurrency behavior appropriate to the external service.
- 05
Test through Automation
Validate and publish the workflow, use a controlled run, and inspect its step input, output and audit history.
Chapter 5 of 10
Ground answers in your documents
A document can be attached to a thread for immediate work, while governed corpora provide reusable organization knowledge with access controls.
| Source | Use it for |
|---|---|
| Thread attachment | A document needed for the current conversation, such as a sample form whose fields should become a Data model. |
| Governed corpus | Reusable internal knowledge that authorized users may retrieve across assistant runs. |
- The assistant can classify, extract, summarize or remove personal data from text.
- Grounded answers include citations back to the available source content; open the citation before relying on a critical claim.
- A person who cannot access the governed source must not gain it merely by asking the assistant; retrieval authorization is checked independently.
- Delete or restrict source content at the governed source, not by relying on a prompt that asks the model to forget it.
Chapter 6 of 10
Worked example: extend invoice intake safely
Use separate preview, draft and release requests so the human operator can inspect each product artifact while live authorization continues to enforce the real boundary.
Read the published Data schema invoice / payables / v1 and the
invoice-approval workflow. Explain:
- the current fields and validation
- every workflow route and external side effect
- the changes needed to capture an optional purchase order number
- the impact on mappings and existing consumers
Do not publish or arm anything. Do not apply changes in this turn.- 01
Review the preview
Compare the explanation with the published Data schema, Automation graph and connected ERP operation. Correct an assumption before any write-capable request.
- 02
Request drafts only
Ask the assistant to create the Data v2 draft, add the optional field, update the Automation draft mapping and run both native validators. Explicitly keep publishing and trigger arming out of this turn.
- 03
Inspect native artifacts
Open the Data draft and Automation validation result. These product screens, and not the chat summary, are the authoritative definitions that would be released.
- 04
Release narrowly
After review, ask for the exact publish and arm operations required. An allowed write may run during this conversation, so verify the new immutable versions and binding immediately afterward.
- 05
Run a controlled example
Create one invoice with purchase_order_number, follow its workflow run and confirm the ERP mapping before broader traffic uses the change.
| Observed outcome | What it means and what to do |
|---|---|
| Preview returned without writes | Review the proposal. A conversational explanation is not a published product artifact. |
| Permission denied | The organization grant, user grant, declared tool or live product role blocked the action. Ask an administrator for the minimum required access; do not try to prompt around it. |
| Data or Automation validation failed | Open the owning product’s validation result, correct the contract or mapping, then validate again before publishing. |
| Provider or tool call failed | Keep the assistant run identifier and failed tool entry. Check whether a partial draft exists before repeating the request. |
| Publish and arm succeeded | Verify the immutable schema/workflow version, armed binding, audit entry and one controlled run in their product screens. |
Chapter 7 of 10
Connect an external AI client through MCP
A compatible AI client can connect to SynaptaGrid’s remote Model Context Protocol server and use the same permission-filtered tool surface.
- 01
Add the remote server URL
Use the SynaptaGrid MCP resource URL shown in Portal. There is no local SynaptaGrid server to install.
- 02
Sign in as yourself
The client discovers the authorization service, registers when supported and opens the normal authorization flow. Do not distribute a shared API key.
- 03
Approve the client
A client that registered itself, such as ChatGPT, shows a consent screen before it gets any access. It names the application, the address its sign-in returns to, what it may do and the workspace it will work in. Check all four before you allow it; cancelling gives the client nothing.
- 04
Inspect the tools
The server exposes only tools from agents explicitly enabled for MCP and allowed by your current grants and product roles.
- 05
Use and review normally
Tool calls use the built-in assistant’s dispatch checks. Review available run details and product audit history for the activity recorded under your signed-in account.
- 06
Disconnect when you are done
Connected applications in your profile lists every client you approved. Disconnecting one revokes the tokens it holds, and connecting it again asks for your approval again.
| Client | Where to add it |
|---|---|
| Claude.ai / Claude Desktop | Settings → Connectors → Add custom connector, then paste the server URL from AI → Connect MCP in your workspace. |
| ChatGPT | Settings → Apps & Connectors → Advanced settings → Developer mode (turn it on first if it is not visible), then Add connector with the same server URL. |
| Claude Code and other native clients | Use the ready-made command or config shown on the same Connect MCP page; OAuth still opens automatically on first connection. |
- Use the server URL shown in Portal → AI → Connect MCP for your deployment, then confirm the connected workspace. Workspace selection and access come from your signed-in connection, not from guessing another host.
- The connection uses MCP Streamable HTTP with protocol negotiation, sessions and server-sent events (SSE). Tools, prompt templates, resources and interactive results depend on what the client supports and what your account can access.
- Connecting a new client does not widen access: the organization and user agent grants, the tools an administrator exposed to MCP, and your live product role all still apply exactly as they do inside the built-in assistant.
- MCP calls retain the server’s permission and applicable approval checks. Approval interactions can differ by client and organization settings; follow the prompt provided for the call.
https://ai-api.synaptagrid.io/v1/egav-ai/mcpClient instructions: https://www.synaptagrid.io/assistant#mcp-skill
Readable instructions: https://www.synaptagrid.io/skills/egav-mcp/SKILL.md
Skill ZIP: https://www.synaptagrid.io/skills/egav-mcp.zip- “Show my open Automation runs that failed in the last 24 hours and explain the first failure.”
- “List the entity types in my Data workspace and show the required fields on the newest one.”
- “Summarize this attached invoice and check whether we already have a matching supplier record.”
- “Draft a supplier-onboarding record type with company identity, tax fields and evidence uploads — do not publish it yet.”
- “Search our knowledge base for how refunds are processed and cite the source.”
- “Look up support ticket SUP-482 and tell me its current status and last update.”
- “Show the audit history for the last change to the invoice-approval workflow.”
- “Which automation workflows call our ERP, and which of them ran in the last week?”
Chapter 8 of 10
Track usage, runs and audit history
A conversation is the user interface; the run and product audit records are the operating evidence.
- Assistant messages link to the run that produced them, preserving the connection between a response and its execution.
- Runs record the agent, caller, state, tool activity and metered usage needed to investigate a failure or unexpected action.
- Product mutations also enter the same audit trail as actions performed through the regular interface, attributed to the signed-in user.
- Credits are reserved before execution, individual runs have a ceiling, and organization administrators can allocate per-user caps and review consumption.
- Organizations may use platform credits, supported bring-your-own provider keys or a self-hosted-only model policy according to their configuration.
Chapter 9 of 10
Know the current safety boundaries
Permission enforcement is real, but model output still requires proportionate human review. Do not describe a prompt instruction as a technical control.
- There is no universal approval gate that pauses every assistant write before execution today. An allowed tool may run during the conversation.
- Ask for a preview and explanation first when a change is consequential, then inspect and test the resulting product artifact.
- The assistant’s workflow diagnosis is read-only, but builder capabilities can change Data and Automation when the user and grants allow them.
- Do not rely on claimed signed AI-run provenance, general run replay, cross-encoder reranking or a customer-facing corpus-management surface; those are not shipped product guarantees.
- MCP tool calls return a result over Streamable HTTP. SSE carries protocol messages; it does not promise streamed model tokens for each tool call.
Chapter 10 of 10
Troubleshoot an assistant run
Separate access, knowledge, tool validation, provider and budget failures before changing the prompt.
| Symptom | Check first |
|---|---|
| Assistant is missing | The early-access feature is enabled for the organization and the user holds the assistant-use permission. |
| A needed action is not available | The agent’s tool list, organization grant, user grant and the user’s role in the owning product. |
| Permission denied during a call | The receiving product’s live permission; cached visibility never overrides the final authorization check. |
| Answer lacks the expected source | The attachment finished processing, the correct corpus is connected and the user may retrieve that content. |
| A generated definition fails | Open the Data schema or Automation validator and correct the reported contract or mapping error there. |
| Run stops for credits | Organization balance, user allocation and the per-run cost ceiling. |
| External API step fails | The registered system’s auth, base URL, imported operation, mapping, throttle and retry state. |
| Remote MCP client cannot connect | Use the exact Portal-provided resource URL, complete sign-in again and verify the client supports remote MCP authorization. |
Continue
Continue with the product guides
Data and Automation are sold separately and work independently. Use both when a workflow should read or change a versioned business record.