Smartsheet is an enterprise work administration platform that a whole bunch of hundreds of organizations depend on. As enterprise groups undertake AI brokers, these brokers want structured entry to the information inside programs like Smartsheet, however most programs aren’t constructed for that. To bridge this hole, Smartsheet constructed a distant Mannequin Context Protocol (MCP) server on AWS that offers AI purchasers direct entry to its knowledge and capabilities. AI assistants like Amazon Fast and Claude Desktop assist customers work together with Smartsheet’s capabilities by way of pure language, analyzing challenge knowledge, updating duties, creating sheets, managing workspaces, and extra.
Enterprises are additionally constructing customized AI brokers for workloads that run with out human prompting. These AI brokers can work autonomously of their roles, coordinating by way of Smartsheet utilizing MCP. A couple of examples are capturing necessities, choosing up duties, attaching take a look at outcomes, drafting documentation. These occur in the identical sheets their human counterparts use, compressing workflows that took weeks into days or hours.
The MCP server connects to Smartsheet’s current APIs and central intelligence layer. It additionally provides an AI-optimized interface on prime, designed to reduce token value, assist stop hallucination, and assist massive language fashions (LLMs) work reliably with enterprise knowledge. Since launch, Smartsheet saved over 3 billion tokens, primarily based on inner telemetry, by way of these optimizations.
On this submit, we cowl a high-level view of the Smartsheet distant MCP structure, with a concentrate on the AWS infrastructure behind it. This consists of safety, governance, scaling and deployment, and the AI-specific optimizations Smartsheet constructed on AWS.
Structure
One MCP layer serves each inner and exterior brokers. Smartsheet’s personal Sensible Help (the in-product AI expertise) and externally linked AI purchasers like Amazon Fast run on the identical infrastructure, with the identical instruments, optimizations, and intelligence stack. This parity is a deliberate architectural alternative: Smartsheet builds as soon as and each agentic consumer advantages instantly.
The architecturally crucial AWS companies within the knowledge path are:
The detailed structure circulate is as follows:
- AI purchasers to API gateway layer to MCP Server: Requests move by way of an API gateway layer (AWS WAF, AWS Protect, AWS Software Load Balancer, and OAuth validation) earlier than reaching the MCP server on AWS Fargate.
- MCP Server to Area Providers: The MCP server calls Smartsheet’s area companies by way of their APIs for transactional operations.
- MCP Server to Intelligence Layer: The MCP server queries the Intelligence Layer constructed on Amazon Neptune and Databricks for cross-project agentic insights.
- Area Providers to Intelligence Layer: Change occasions stream by way of Amazon Kinesis and Apache Flink into the Amazon S3-backed intelligence layer. The Intelligence Layer follows the medallion structure.
Determine 1: Smartsheet MCP Server structure on AWS.
Supporting companies corresponding to edge safety, container registry, observability, secrets and techniques seem within the related sections within the following sections.
Deployment and scaling
On the scaling facet, AI visitors differs from typical request patterns. Brokers autonomously orchestrate sequences of software calls, firing a number of requests in a second as they work by way of a activity, then going quiet whereas the mannequin causes. This bursty sample calls for a scaling technique that responds to each sudden spikes and sustained throughput.
To deal with and validate this sample, Smartsheet constructed the MCP server to run on AWS Fargate for Amazon ECS. ECS Auto Scaling makes use of target-tracking insurance policies, combining visitors quantity with compute utilization. Compute-aware scaling issues right here as a result of every request entails server-side processing like LLM-optimized serialization, not solely proxying. Intensive load testing beneath production-like visitors patterns validated that the infrastructure absorbs agent bursts with out degradation.
On the deployment facet, delivery updates with out disrupting lively agent periods is equally crucial. Container pictures are saved in Amazon Elastic Container Registry (Amazon ECR) and rolled out by the continual integration and steady supply (CI/CD) pipeline by way of a layered security internet. ECS deployment circuit breakers detect failing containers throughout rollout and routinely revert to the final steady model. Handbook intervention just isn’t wanted, and buyer impression is prevented.
Deployments roll out to the smallest area first, following the AWS Properly-Architected precept of decreasing impression radius. After every area, automated end-to-end checks validate software conduct towards the dwell surroundings. Canary checks run each quarter-hour, executing a multi-step MCP workflow by way of the complete authentication and gateway path. Outcomes feed into the monitoring stack so degradation surfaces earlier than buyer studies. The ECS Fargate with ALB sample is documented within the AWS Steering for Deploying MCP Servers.
Governance and observability
For enterprise prospects, governance is the gating issue for AI adoption. Smartsheet constructed it into the software framework itself: entry management, error dealing with, and audit trails ship with each software by default.
Entry is tiered per group: directors can activate AI entry globally, limit to non-destructive operations solely, or open up full write and damaging capabilities, giving every group management over their adoption curve. Instruments carry MCP protocol annotations like readOnlyHint and destructiveHint so AI purchasers apply acceptable affirmation flows routinely.
The server emits OpenTelemetry indicators (logs, traces, and metrics) throughout the complete request lifecycle. Each software invocation captures the utmost context potential inside privateness constraints: person, group, software identify, final result, and extra. This offers the inspiration for utilization insights and compliance auditing.
Agent visitors is more durable to watch than conventional API visitors. A single person request can produce a series of software calls, and failures typically hint again a number of steps. Smartsheet is extending its observability with agent-first id and tracing, correlating context throughout software chains. Logs stream by way of Amazon Kinesis into Amazon OpenSearch Service following the AWS Observability Greatest Practices sample, with infrastructure metrics surfaced by way of Amazon CloudWatch. Datadog offers per-tool software efficiency monitoring (APM) visibility, and PagerDuty handles incident routing.
Each invocation additionally emits a structured analytics occasion by way of Amazon Easy Queue Service (Amazon SQS) into the Intelligence Layer. This closes the suggestions loop: manufacturing utilization knowledge informs which instruments to prioritize and the way optimization methods carry out throughout actual workloads.
Securing AI agent visitors
The MCP server runs behind the identical safety infrastructure as Smartsheet’s manufacturing APIs. AWS WAF and AWS Protect are on the edge, with personal subnets in a digital personal cloud (VPC), mutual TLS (mTLS) for service-to-service calls, and an OAuth2 proxy that rejects unauthenticated requests earlier than they attain compute. The MCP server follows the defense-in-depth mannequin within the AWS Steering for Deploying MCP Servers. The API gateway layer handles authentication and scope validation. Area companies deal with fine-grained permissions. If a person can’t entry a sheet by way of the UI, they will’t entry it by way of MCP both.
AI visitors provides a particular rate-limiting problem. A single person query can set off a number of software calls in seconds. Many enterprise customers sit behind shared company proxies making IP-based charge limiting unreliable.
To handle this, Smartsheet carried out layered charge limiting by way of AWS WAF. Three layers work collectively: blanket safety on the periphery, per-user metering utilizing customized aggregation keys on an id header, and path-specific controls for costly operations. Per-user metering means periods are metered individually somewhat than pooled by IP. The layered charge limiting follows the three most necessary AWS WAF rate-based guidelines sample.
Testing non-deterministic AI workflows
Smartsheet maintains the usual testing layers: unit checks, integration checks, tool-level validation. The MCP server, nonetheless, introduces a testing problem that conventional API companies don’t face. A traditional API response will get rendered deterministically by the UI. An MCP software response passes by way of an LLM first. The mannequin interprets it, causes over it, and generates what the person truly sees. That layer of non-determinism modifications what “appropriate” means for testing.
Smartsheet invests closely in end-to-end workflow checks that embody the LLM within the loop. These checks simulate reasonable enterprise situations: creating workspaces, writing knowledge, querying outcomes, and verifying that the mannequin’s interpretation is smart to the tip person. These checks run within the CI/CD pipeline (GitLab CI with runners hosted on AWS) and repeatedly as canary checks towards every manufacturing AWS Area.
Optimizing for AI consumption
As enterprises scale AI agent deployments, token consumption turns into an actual value driver. Each software response prices cash on the LLM and competes for context window capability. Most MCP software calls at the moment run with out sub-agent orchestration. The agent calls instruments straight, one by one, reasoning between every step. With out clever software design, this will get sluggish, costly, and error-prone quick. Every software name should be self-contained and environment friendly by itself, which is why Smartsheet optimizes at three ranges:
- Progressive disclosure that caps token consumption per response.
- Strongly typed software schemas that assist stop hallucinated parameters and wasted calls.
- A proprietary serialization format that reduces token depend by 35–47 % on data-heavy responses.
Progressive disclosure
Every software response targets a token finances. The server dynamically calculates what number of rows match primarily based on column depend and knowledge density. For instance, a sheet with 5 columns can return extra rows than one with 15 columns, however the complete stays inside finances. Whether or not a sheet has 50 rows or 50,000, the response measurement stays bounded. The mannequin sees sufficient to orient, then narrows with filters primarily based on what the person is definitely asking about.
Metadata fields inform the mannequin precisely what occurred: is_sampled signifies whether or not knowledge was truncated, rows_in_sheet provides the complete depend, rows_actual reveals what number of had been returned, and filters_applied describes lively filters. The mannequin makes use of this to determine whether or not it has the complete image or must slim its question with filters. Progressive disclosure is a server-side choice. The MCP server handles budgeting and sampling, whereas the metadata it returns provides the AI consumer the indicators to orchestrate follow-up queries by itself.
Determine 2: Progressive disclosure in motion: the AI consumer receives sampled knowledge with metadata, then makes focused follow-up requests.
Preserving LLMs grounded: Schema-driven software contracts
Grounding the LLM is crucial. With out constraints, fashions hallucinate parameter names, invent operators, and waste tokens on failed calls. Every software publishes a strict JSON Schema by way of MCP’s software discovery, generated from Pydantic fashions. Parameters are constrained to legitimate enums, column names are validated towards the precise sheet earlier than execution, and mismatches return structured errors with legitimate choices as a substitute of failing silently. Schema validation catches hallucinations on the boundary and means brokers navigate the software catalog reliably with out trial and error.
Token-efficient serialization
JSON’s structural overhead (braces, quotes, repeated keys) usually consumes 15–25 % of a response’s token depend. For a server returning spreadsheet knowledge with hundreds of rows, the overhead provides up quick.
Smartsheet constructed a proprietary serialization format that alleviates this overhead. Key names seem as soon as as a substitute of repeating per row, and structural syntax is changed by delimiters that tokenize extra effectively. On a consultant 33-item filtered question, the optimized response is roughly 3,900 tokens versus roughly over 6,000 tokens for the equal JSON, roughly 35 % fewer tokens carrying the identical data. At 1,000 rows the hole widens additional, as a result of JSON repeats key names on each object whereas the optimized format declares them as soon as.
What’s subsequent for Smartsheet
AI brokers combine with Smartsheet by way of MCP at the moment. Within the first 4 weeks after common availability (GA), Smartsheet noticed over 87 % week-over-week person progress.
MCP is the distribution layer. What comes subsequent is intelligence on the connection level itself. One instance is assets that form themselves to the particular person, crew, and organizations utilizing them. One other is brokers that run autonomously on workflows, and a routing layer that lets specialists hand off reasoning to one another somewhat than beginning chilly on each step. Similar MCP connection, completely different intelligence per buyer, with no deployment required.
AWS is evolving its infrastructure to satisfy these rising agentic necessities. Amazon Bedrock AgentCore offers runtime execution, discovery, personalization, and governance by default. Smartsheet continues to undertake and form these capabilities with AWS.
The MCP protocol itself continues to evolve. Elicitations enable human-in-the-loop affirmation earlier than damaging actions. MCP Apps deliver interactive UI straight into AI conversations. Duties assist asynchronous, long-running operations. Smartsheet is evaluating these as they mature.
AI strikes quick. Constructing on AWS provides us the infrastructure to maintain tempo whether or not which means new protocols, new fashions, or totally new agent architectures.
To connect with Smartsheet’s MCP server, go to the AWS Market itemizing, or see the Smartsheet MCP documentation.
In regards to the authors
