MCP 2025-11-25 JSON-RPC 2.0 design.md API keys # Build with Loop Set up an agent with clicks, connect the tools you already use, or build directly against our API. However you work, Loop gives every go-to-market conversation your brand and business context. [Meet with a GTM Engineer](https://loopllm.ai/contact) Rather not build it alone? Our specialists will set it up with your team, from a two-week coaching engagement to a monthly managed service. [See Services & Accelerators](https://loopllm.ai/accelerator) Four ways to build ## How do you build? From setting up an agent with clicks to scripting against the API, pick the option that matches how you work. Hover any technical term for a plain-English definition. Configure with clicks, not code Author templates, skills, and collections in the editor, and adapt the agents you already have. Bring context to every conversation Loop loads your context, brand, and voice into every conversation automatically. Connect your tools and AI Built-in integrations, import from any web address, and your context inside Claude Desktop or Cursor over MCP. Integrate the Loop API to anything API keys and the MCP JSON-RPC API to build custom automations on Loop. ## Configure with clicks, not code Create the building blocks in the editor. Each one is reusable, built on your own organization’s content, and available right away to chat, automations, and any connected client. Editor Write docs, knowledge, and skills in a rich editor with mentions and variables. Template Architect Turn any doc into a reusable template with {{variables}} and form fields. Skills Package repeatable instructions an agent loads on demand, authored as plain markdown. Collections Bundle the knowledge your agents and skills retrieve from. Adapt agents Fork a system agent and reshape its knowledge, voice, memory, and granted tools. Forms Add fields, validation, and capture to any document without touching a separate builder. Adapt the agents you have on any plan. Create custom agents from scratch on Enterprise . 2 · The Context Engineer ## Bring context to every conversation Loop keeps five things loaded into every conversation automatically: your company info, brand, voice, memory, and house rules. You don’t have to paste any of it in yourself. Company context Your mission, industry, and values. Brand Colors, identity, and your design.md. Voice & writing Formality, warmth, and tone dimensions. Memory Preferences and facts learned over time. Rules Never-say, emphasis, and guardrails. #### The variable cascade Every prompt is filled in from five levels before it reaches the model, in this order: document override, user preference, resolver config, org default, system default. The most specific level wins. ##### Namespaces `env``user``org``file``contact``company``deal``input``calc` Bare `{{field}}` uses smart resolution across all namespaces — A variable prefix (org, user, contact, company, …) that declares where a value comes from. . handlebars ``` Write a cold email to {{contact.firstName}} at {{company.name}} on behalf of {{org.name}}. Tone: {{user.formality}} · {{user.warmth}} ``` text ``` // compiled before it hits the model → Write a cold email to Sarah at Acme Corp on behalf of OpenBrain. Tone: professional · warm ``` 3 · The Connector ## Connect your tools and AI Bring the outside world into Loop, and bring Loop into the tools you already use. #### Native integrations you deploy, not build Google Workspace Team+ Slack Team+ Notion Coming soon Composio · 100+ apps Coming soon Twilio · SMS/WhatsApp Coming soon Web Import Paste any URL and Loop reads the page straight into your context. Pro+ oEmbed — An open standard that lets one URL render as rich, inline content on any platform that supports it. Drop a Loop doc link into Notion, Confluence, or Slack, and it renders inline. Free+ #### Your context in your AI client MCP · Enterprise · Beta Loop runs the MCP — Model Context Protocol: the open JSON-RPC standard an AI client speaks to reach your agents, tools, and documents. server. Paste the config it gives you into Claude Desktop, Cursor, or Windsurf, and your agents, documents, and brand voice are right there. No code. json ``` // claude_desktop_config.json (paste & go) { "mcpServers": { "loop": { "url": "https://app.loopllm.ai/api/mcp", "headers": { "Authorization": "Bearer loop_your_api_key" } } } } ``` Want to write code against MCP instead of just deploying it? That’s [the Integrator →](https://loopllm.ai/builders#integrator) 4 · The Integrator ## Integrate the Loop API to anything #### API keys All API requests use Bearer token auth. Keys are prefixed `loop_` and carry read and write access to your organization, which covers every method below. Requires an Enterprise plan. `tools:execute`Call MCP tools and invoke actions `resources:read`List and read documents, collections `agents:read`Access agent definitions Owner or Admin only Keys are shown once at creation and stored as a SHA-256 hash. Lost your key? Revoke it and generate a new one. bash ``` curl https://app.loopllm.ai/api/mcp \ -H "Authorization: Bearer loop_your_api_key" \ -H "Content-Type: application/json" \ -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}' ``` MCP API methods `initialize` Negotiate protocol version `tools/list` List available Loop tools `tools:execute` `tools/call` Invoke an agent or tool by name `tools:execute` `resources/read` Read a document or collection by URI `resources:read` Decision & evaluation models ## Small models make the calls, so large models can write A decision model writes no text. It picks one of a set of options, places something on a scale, or says how likely a statement is to be true, and reports how sure it is. Loop uses them for three jobs. Routing Which agent answers a chat message, which skill fits the request, which tools are offered first, and which sources the agent reads. Classification Whether a request asks for a piece of work and which template fits it, and which of a skill’s questions the message already answers. Evaluation Each chat answer scored 0–100 on relevance, usefulness, accuracy, brand, context, and clarity. DeepEval checks run beside it: faithfulness to the sources read, answer relevancy, toxicity, and across the conversation, turn relevancy, knowledge retention, and role adherence. Models Jev · default Jev 1.13 Jev 1.14 GPT-6 Luna Decisions Liquid d1 Microsoft Decision-1 An owner or admin picks the model in Settings → AI, or has your regular chat model answer instead. If a decision model fails, the same questions go once to your regular chat model. Every decision is logged with its answers and how sure it was, and decisions never decide who may see or change what. Open standards ## Built on open standards, so nothing is locked in Loop supports the open standards AI tools are settling on, so your context, skills, and brand stay portable and readable instead of locked inside one product. MCP JSON-RPC 2.0 · 2025-11-25 The open Model Context Protocol lets any MCP client reach your agents, tools, and documents. skill.md Markdown + frontmatter Skills are portable markdown files: readable, versionable, and embedded into retrieval. design.md /api/orgs/[slug]/design.md Your whole brand as one markdown file, served to any tool over an API key. llms.txt Site-level discovery Publish your public content as a machine-readable map, so AI crawlers and agents can find and cite it. oEmbed GET /api/oembed Public Loop documents embed inline anywhere oEmbed is supported. Talk to us ## Start building now Tell us what you’re building, and we’ll help you find the right path: from clicks to connections to custom code. [Contact us](https://loopllm.ai/contact) --- Source: https://loopllm.ai/builders