Product Updates

Introducing Gram: Build MCP servers that perform

Sagar Batchu

September 9, 2025 - 8 min read

---

MCP, MCP everywhere, nor any tool to use

In just a year, MCP has gone from protocol white paper to critical infrastructure. As developers scramble to get a handle on this new technology, MCP servers are everywhere, and yet, good MCP servers remain elusive.

Any developer can create a basic MCP server in minutes. The code isn’t complicated and the infra is straightforward. But building a server that AI agents can actually use effectively is a nuanced, ever-evolving challenge. Most MCP servers fail for one of four reasons:

1. They expose too many tools, creating decision paralysis and tool confusion for LLMs.
2. Tools lack the rich descriptions and examples that help LLMs understand when and why to use each tool.
3. MCP servers provide CRUD-based tools lacking context, which forces LLMs and agents to construct workflows on the fly.
4. They have an incomplete authentication story preventing production usage.

The result? Servers that may “technically” be MCP-compliant… but in practice, confuse agents and produce unreliable results. Ultimately, the promise of MCP and agents goes unfulfilled, and dev teams give up.

Gram: build tools that perform

Gram is an open source platform that bridges the gap between tools and AI agents. Our unique focus is tool design. We help you curate and compose your APIs into intelligent custom-built tools that agents can actually use effectively.

Most frameworks for building MCP servers focus on the mechanics of server creation: decorators, custom functions, and infrastructure. While these are important foundational elements, the critical factor for agent success is tool design: how you structure, describe, and organize the capabilities your server exposes. LLMs don’t interact with your code, they interact with your tool definitions, descriptions, and prompts.

Gram enables teams to add context, refine prompts, and compose custom tools until your APIs can be consumed effectively by LLMs.

Here’s how it works.

1. Curate Toolsets → Eliminate AI Confusion

When you upload your API to Gram, we convert your API endpoints into primitive tools. We help you identify which capabilities actually matter for your use cases and filter out the noise. A 600-endpoint API becomes a focused 5-30 tool collection that AI agents can navigate confidently.

All your tools are grouped into toolsets. You can remix tools across different APIs into a single toolset ensuring you have all the right tools for a specific use case rather than a single API.

2. Add Context → Improve AI Performance

APIs designed for developers often lack the rich context that AI agents need. Our platform lets you enhance tool descriptions, add business logic, and provide examples that help LLMs understand not just what each tool does, but when and why to use it.

Every time you update a tool, immediately test it out in a convenient playground.

3. Define Custom Tools → Create Complete Solutions

This is where Gram truly differentiates itself. Instead of forcing AI agents to orchestrate multiple tool calls together manually, you can define what we call “custom tools,” tools that chain together smaller atomic tools and represent specific business operations.

For example, let’s say you ask your AI a question like “Summarize ACME Corp’s health”. Without custom tools, your AI might need to figure out, from a plethora of tools available, that it needs to (a) find the ACME Corp customer record and id; (b) search for the id within the CSAT scores table; (c) search for ACME Corp using a specific tool from your CRM; (d) retrieve relevant notes within your CRM; (e) summarize all the above.

As the number of tool calls that the AI needs to independently figure out increases, the chance of making an error increases: e.g. if there’s a 5% chance of each individual call failing i.e. 95% chance of success, then across five calls that figure drops to (1-5%)^5 = 77% chance of success.

With custom tools, you can instead create a tool that is targeted at a specific use case, and that calls specific tools that you define — decreasing the chance that AI encounters tool confusion.

Prototype fast, scale faster

Getting started with Gram takes under a minute, but our platform is built to grow into a complete MCP control plane that enterprises need.

OAuth 2.1 Compliance: MCP servers deployed through Gram can optionally include our OAuth 2.1 proxy with Dynamic Client Registration support, and PKCE flows. Your tools are secure by default, not as an afterthought. Already have your own OAuth flow implemented? Bring your own OAuth authorization server and add it in front of any Gram hosted MCP Server with just a few clicks.

Centralized Management: You get unified control across all the agent tools within your org, with role-based access control, comprehensive audit logging, and compliance reporting. It’s the difference between scattered tools and a strategic tool repository.

Production Infrastructure: Our hosted infrastructure means automatic scaling, uptime SLAs, zero downtime deployments, and 24/7 monitoring. No local setup, no infrastructure management, no maintenance overhead.

Every toolset in Gram is immediately available as a hosted MCP server. Whether your teams use Claude Desktop, ChatGPT, Cursor, or any other MCP-compatible platform, integration is instant and tested for different AI ecosystems.

Real Impact: From Prototype to Production

MCP servers have been transformational for the companies we work with. We see three primary use cases driving adoption:

1. Launching Public MCP Servers

Companies are exposing their APIs as MCP servers to make their platforms AI-native. By creating well-designed MCP servers, businesses enable AI agents to seamlessly integrate with their services, opening up new distribution channels and user experiences. Instead of users needing to learn complex APIs or navigate web interfaces, they can simply ask an AI agent to perform tasks using natural language. We want to make launching mcp.yourcompany.com as easy as possible.

2. Embedding AI-Native Experiences

Product teams are embedding AI capabilities directly into their applications using MCP servers as the integration layer. All major agentic frameworks like Langchain, OpenAI Agents SDK, and PydanticAI support MCP servers for accessing tools. MCP servers interact with their existing APIs and data sources allowing developers to embed AI chat, image, and video generation, and audio synthesis into their applications. This approach dramatically reduces development time while providing users with intuitive, natural language interfaces.

3. Internal Workflow Orchestration

Every company has internal admin APIs for operational capabilities: user management, billing operations, system diagnostics, data analytics, configuration controls, etc. Teams are replacing internal dashboards, saved SQL commands, and shared bash scripts with AI agents powered by MCP servers. Complex multi-step operations that previously required navigating multiple systems can now be accomplished with a simple natural language request.

Why we open sourced Gram

We made the decision to open source Gram  because the MCP ecosystem is evolving rapidly, and we realized that much of what we’re building will continue to change in the coming months. MCP has already transformed from a local-only tool to supporting remote servers, and this pace of innovation shows no signs of slowing.

With an open source approach, we can engage directly with the community as we build this new piece of the AI engineering stack. This collaboration is essential for integrating with other agentic frameworks and ensuring Gram works seamlessly across the broader ecosystem.

Most importantly, Gram is infrastructure that we’re asking companies to rely on. Being able to see the code, understand how we build it, and evaluate our engineering practices gives users confidence in what we’re doing. Transparency builds trust, especially for critical infrastructure.

The tech stack includes:

• TypeScript  – dashboard language.
• Golang  - backend language
• Goa  - design-first API framework in Golang.
• Temporal  - workflow engine.
• Polar  - usage-based billing.
• Speakeasy  - generated SDKs.
• Our own homegrown OpenAPI parser 

Contributing: The entire platform is open source with a comprehensive development guide. Getting started locally requires just running a mise script, with only two external dependencies: a Temporal key and OpenRouter key. We’re particularly looking for contributions around MCP server capabilities, expanding our OpenAPI parser to support more API shapes, and implementing newer features from the MCP specification.

Getting Started

As of today, Gram is in public beta and anyone can sign up and get started.

Pricing is usage-based with a generous free tier (1K tool calls) to encourage teams to experiment and iterate quickly.

The future belongs to organizations that make their capabilities easily accessible to AI agents. The question isn’t whether your company will need MCP—it’s whether you’ll lead the transformation or follow it.

Try it out today Join Gram  and start powering AI agents by leveraging your APIs.

What’s coming

Looking ahead, we’re expanding the platform in three key directions:

Gram Functions - Moving beyond OpenAPI-only tool creation, teams will soon be able to upload custom TypeScript and Python code to create sophisticated agent tools that don’t require existing APIs. This opens Gram to any team building AI workflows, not just those with existing OpenAPI specifications.

MCP server import - Rather than building yet another marketplace, we’re enabling teams to import and manage the best MCP servers from across the ecosystem—including official servers from Anthropic’s registry, official first party servers from the likes of Notion, Linear etc. and third-party marketplaces like Smithery. Your Gram workspace becomes a unified control plane for all your organization’s MCP servers whether you’re making a server or consuming an external one.

Self-hosted data plane - Enterprise teams will gain self-hosted data plane options to keep API traffic within their VPC, comprehensive observability and audit trails, role-based access controls, and compliance certifications including SOC2 Type 2. Plus, embeddable Gram Elements will let you add curated chat experiences directly into your applications.

Check out the complete roadmap here .

Interested in learning more about our approach to MCP and enterprise AI integration? Book time with our team