Connect an MCP client to unlob

The two connection forms

Almost every MCP client accepts one of these.

Remote HTTP — the hosted service. Nothing to install.

url:     https://api.unlob.com/mcp?profile=grounding
header:  x-api-key: ulb_…

?profile=grounding lists ground and get_document, which is the right surface for most tasks. Drop it to list every tool — see Profiles.

stdio bridge — for clients that only speak stdio. mcp-remote is a small proxy that speaks stdio to your client and HTTP to us:

npx -y mcp-remote https://api.unlob.com/mcp --header "x-api-key: ${UNLOB_API_KEY}"

Prefer the remote form where the client supports it: it is one fewer process, one fewer thing to keep updated, and it does not need Node installed.

Clients

Client	Native remote HTTP	Page
Claude Code	Yes	Claude Code
Claude Desktop	Via connectors	Claude Desktop
Cursor	Yes	Cursor
VS Code (Copilot)	Yes	VS Code
Codex CLI	Version-dependent	Codex CLI
Cline, Windsurf	Version-dependent	Cline and Windsurf
ChatGPT	Via connectors	ChatGPT
MCP Inspector	Yes	MCP Inspector

Building your own client, or using a framework? Seven framework guides start at LangChain, and n8n covers the no-code path. Or call the REST API directly.

Keep the key out of the file

Every configuration below can hold a literal key, and most of these files end up in a repository or a dotfiles backup. Use an environment variable where the client expands one, and check that the file is git-ignored where it does not.

ulb_ keys are shown once and can be rotated freely, so if one does end up committed, rotate it in the console rather than trying to scrub history.

After connecting

Ask the client to list its tools. On the grounding profile you should see two:

ground  get_document

On the default profile, every tool:

web_search  get_document  similar  why_not  browse  related  corroborate
authorities  dossier  path  assemble_context  ground  evidence_changes

Zero tools and no error message is the characteristic MCP failure, and it is almost always the handshake rather than the tools. Troubleshooting has the one curl command that tells you which step failed.

Telling the agent what it has

Most clients read the server’s instructions, so the agent already knows roughly what unlob is. Two habits are worth adding to your own prompt, because they are what make the difference between using this API and using any search API:

• Prefer ground for “what should I know about X, and can I trust it”; prefer web_search for “find the page that says Y”.
• Read status and coverage on every ground answer before using the evidence, and never restate a partial answer as complete.

Or skip the prompt: install the Agent Skill, which says all of this in the shape a coding agent reads on its own.