Skip to main content
SynaptaGrid Platform
← All capabilities
Available now

Platform capability · AI Assistant & Agents

Connect a registered app to AI agents and MCP

Agents and tools act with the access of the signed-in user

A registered app declares its AI providers, agents, tools, MCP servers, document collections for search, memory scopes and evaluation datasets in the AI section of its app declaration. Publishing the app declaration copies that catalog into the AI service. A tool call uses the signed-in user’s token, so the service that receives it applies its normal permission rules. Agents can run in the platform chat, or through the OAuth-protected MCP interface used by external AI clients and code editors.

Built in

What you get

  • Registered apps declare agents, tools, MCP servers, document collections, memory scopes and evaluation datasets in the app declaration
  • A tool call runs as the signed-in user, and the receiving service checks every action. There is no AI service account with extra rights
  • An agent that hands work to another agent can use only the tools both are allowed, so handing over cannot add access
  • Agents with defined input and output, checked against that definition, and free-form chat agents
  • Memory for each agent is limited to a thread, an organization or a user, and an agent cannot read outside its scope
  • Search that combines keyword matching with meaning-based search
  • Search data is stored in PostgreSQL with the pgvector extension, and runs the same locally and on AWS
  • Embeddings for meaning-based search come from a provider you configure
  • Works with Anthropic, OpenAI or your own model endpoint, with one price list for all providers
  • Answers from search cite their source documents
  • A credit pool for each organization, with included and purchased credits. Administrators set each user’s share: unlimited, a percentage, a fixed amount or an equal split
  • Use any agent from chat or over MCP, including external AI clients and code editors. MCP access uses OAuth 2.1 and a token for each user
  • Customers connect Claude or ChatGPT to their own workspace through the MCP connector, with a consent screen, and work with their records, workflows and reports under their own permissions
  • Sensitive tool calls are built to wait for a person to approve them. Support staff cannot approve on a customer’s behalf
  • Agents take real actions through declared tools, and every call passes the same sign-in, permission, feature and quota checks
  • Chat threads with saved history, streamed answers and a record of cost, response time and tool calls for each turn
  • Operators build agents in the operator console and turn them on or off for each plan
  • Connect external MCP servers. Their tools stay off until an operator reviews and enables them
Under the hood

How it works

Declare your agents and tools, and publishing registers them

A registered service declares agents, providers, tools, MCP servers, document collections, memory scopes and evaluation datasets in the AI section of its app declaration. Publishing the app declaration synchronizes that catalog into the AI service. Because the catalog is part of the declaration file, it is under version control with the rest of the app. An HTTP tool also needs a definition of its input and an authorized customer-facing endpoint in the service that owns it.

Tool calls use the signed-in user’s permissions

A tool call goes to the customer-facing API of the service that owns the tool, with the signed-in user’s token. That service checks the user’s role, plan feature and quota, and the audit trail records that user as the one who acted. When an agent hands work to another agent, the second agent can use only the tools that both are allowed, so handing over cannot add to the first agent’s declared tools.

Keyword search and meaning-based search together

Keyword search (BM25) finds exact terms such as error codes and identifiers. Vector search finds text with a similar meaning when the question and the source use different words. Reciprocal Rank Fusion merges the two ranked lists without a fixed weight for either.

Credit use is counted for each organization

Each run records the provider’s usage and takes the calculated charge from the organization’s credit pool. Credits included in the plan are used before purchased packs. Authorized administrators can buy packs and set each user’s limit as unlimited, a percentage, a fixed amount or an equal split. When a user reaches the limit, the assistant replies with a clear credit-limit message.

Usage, cost and tool calls for every run

Every assistant turn shows the model used, the response time, the number of tool calls and the credits charged. The record of a run holds every tool call with its inputs, results, response time and effects, so support can trace an answer back to the exact search and tool inputs that produced it. Sampling of production runs, A/B tests of prompt versions and scheduled evaluation runs for each agent are planned.

Further additions

Additions that extend the capability described above.

  • Scheduled evaluation runs and A/B tests of prompt versions
Pairs well with

Related capabilities

App SDK

Register features, events, permissions, AI tools and infrastructure needs in one app declaration

Authorization

Roles, attribute rules and resource hierarchies in one permission system

Audit

An audit trail that can only be added to, where tampering shows

Observability

OpenTelemetry, traces, logs and dashboards

Want a deeper walkthrough of AI Assistant & Agents?

A solution architect will demo the live service against your domain model.

Get a deployment assessment