Munich · Est. 2024

Vendor agnostic
control plane
for enterprise AI agents.

Bevel is a vendor agnostic control plane where your company's AI agents, context, skills, tools and permissions are defined as files your company owns in its own infrastructure, served to any agent runtime over MCP.

Talk to us How it works

Four things fully specify an agent.

What it knows

Context

Typed knowledge nodes with provenance on every fact: where it came from, who last touched it, when it was verified. Compiled into a graph you can traverse, mass-update and build dashboards on.

How it works

Skills

Procedures written in plain Markdown, not prompt fragments buried in a config. Readable by the people who own the process, reviewable in a diff, portable to any runtime.

What it may do

Tools & permissions

Tool manifests with secrets held in a vault, and access rules that say which agent may read which file and call which endpoint. Changes go through review, like everything else.

Who it acts as

Identity

Each agent is a named actor with its own credentials and its own scope, not a shared service account. Every action it takes is attributable to it.

A company whose agents live inside a vendor's product doesn't own its agents.

Today context, tools, skills and permissions live inside a vendor's harness: instructions in the vendor's config, tool wiring in per-vendor connector panels, knowledge in per-vendor grounding stores, and the hard-won lessons scattered across chat logs nobody reads and no other agent can inherit. Switchover is extremely costly.

Where it lives today

Where it lives with Bevel (all within your infrastructure)

ContextPer-vendor grounding stores, re-uploaded per tool

ContextKnowledge nodes with provenance, in your repo

SkillsSystem prompts inside each vendor's console

SkillsWritten procedures in Markdown, reviewed like code

ToolsConnector panels, wired once per vendor

ToolsTool manifests, declared once, served to every runtime

PermissionsWhatever the vendor's admin UI exposes

PermissionsFile-level access rules under change request

IdentityShared service accounts and pasted keys

IdentityPer-agent credentials, secrets held in a vault

LearningsChat transcripts nobody reads

LearningsReviewed edits to the artifact everyone shares

Your repo is the source of truth. Every runtime reads from it.

Your repository

your-company/
├─ knowledge/
│  ├─ customers/…
│  └─ processes/…
├─ skills/
│  ├─ qualify-supplier.md
│  └─ draft-rfi-response.md
├─ tools/
│  ├─ salesforce.yaml
│  └─ sharepoint.yaml
├─ agents/
│  └─ tender-desk.yaml
└─ access/
   └─ policy.yaml

Plain Markdown and YAML. Branches, change requests and diffs are git's, not features we rebuilt.

MCP · UTCP

Any agent runtime

Claude Codedesktop

Cursordesktop

ChatGPTdesktop

opencodedesktop

Background agentsserver-side

In-platform agentself-hosted models

Agents connect and read exactly what they're permitted to read. One-click setup for the tools your employees already use, and the same governed surface for anything running unattended.

Working with

UNITE EGYM LUMINOVO WORKPATH AVI MEDICAL

Supported by

META NVIDIA GOOGLE CDTM UNTERNEHMERTUM NETLIGHT EXIST

What our customers say.

Bevel is our trusted partner for our ongoing AI transformation. On top of their platform that unifies and validates a single source of versioned truth for all people and their AI agents across the company, they support us in building AI use cases and reinventing processes: agents for RFIs, tender management, and GTM, to name a few. Bevel maintains the reusable context layer as we transform the company, helping us unlock business value through concrete use cases and enabling our people along the way.

Bevel has been working with our GTM team to make us more AI-native. They unified fragmented data across Salesforce, Meta, and Google and built campaign agents that let our team design and run GTM campaigns more efficiently, with much less manual work, and they continue to support and maintain what they've built. They don't just have genuine technical depth in AI; they also have a real understanding of how to apply it for business value. They got into the weeds, dug into our workflows to figure out where AI unlocks value, and then built it. Bevel is reliable and consistently delivers, and I'd recommend them to any company looking to become more AI-native.

Bevel brings deep technical expertise, from their experience working at the AI protocol level and in open source. They built us a sophisticated market intelligence agent that bridges scattered external market data with internal company signals and delivers it to the right people, a genuinely complex problem they solved well. On top of the technical depth, they were fast and delivered exactly what they promised. I'd highly recommend them to any company that needs serious engineering behind their AI transformation, not just surface-level implementations.

Bevel has been a trusted partner for us. They co-built with us a first-of-its-kind AI agent that works with complex data of millions of electronics parts and makes them accessible for electronics procurement professionals. They combine genuine AI expertise with real technical depth, and they consistently deliver, fast. I'd recommend Bevel to any company undergoing an AI transformation.

Built by the team behind UTCP and the first ever library for Code Mode (Code Execution for Tool Calling).

We build at the protocol layer this depends on.

Our team created the Universal Tool Calling Protocol, an open alternative to Anthropic's Model Context Protocol, championed by engineers at AWS, Adobe, Royal Bank of Canada and MuleSoft. It is what lets every capability in Bevel be exposed as a plain HTTP endpoint, with no proxy server in the middle. Our CodeMode library is what lets agents chain those tools through code instead of one call at a time.

CodeMode stars: 1.5k+
CodeMode downloads: 45k+
Championed at: AWS · Adobe · RBC

Specification

utcp.io

The protocol, its docs, and how to describe tools you already have.

Source

GitHub

Reference implementations and SDKs, open source.

Library

CodeMode

Tool-chaining through code execution rather than one call at a time.

Thank you very much for building UTCP. It's incredibly helpful for multi-tenant agentic AI applications, where MCP isn't always the best fit. I was actually reverting MCP servers back to hardcoded tool descriptions for my agentic AI systems when I came across UTCP, and I think it's fantastic to finally have a universal standard for this. I fully support UTCP, as I see big value in it for building scalable, multi-tenant AI applications.

Adobe

Andrew Anokhin · Sr. AI / ML / Data Architect

UTCP is a great addition to the agent-tooling space as an alternative to MCP. What I appreciate about UTCP is that it doesn't ask you to rebuild anything. You describe the tools you already have, and agents call them natively, with no proxy server in the middle, no extra hop to operate and secure. It turns “integrate AI with our APIs” from a project into a description.

Amazon Web Services

Luka Perrozzi · Senior Solutions Architect

Git-backed control plane for enterprise AI agents

Tell us what your organization needs.