Connect generates everything an agent needs to use a latch, for the latch you select. You should never be pasting a token into a config file by hand.
What it gives you
- MCP configuration for Claude Code, Claude Desktop, Cursor, and OpenClaw, with the file path each one wants.
- Environment variables:
LATCH_URL,LATCH_TOKEN. - An agent prompt describing what this specific latch can do: its methods, its endpoints, its rate limits, its mounts, its budget. Drop it into the agent's instructions and it will stop guessing.
- A capability preview: what the agent will and will not be able to do.
- A test console to exercise the latch before you hand it over.
The tools an agent gets
Over MCP, a Latch-connected agent has:
| Tool | What it does |
|---|---|
latch_capabilities |
"What am I allowed to do?" Returns this latch's scope. |
latch_authorize |
Checks policy and, when allowed, executes the request server-side without releasing the credential. |
latch_proxy |
Makes a raw governed HTTP request. |
latch_mcp_tools |
Initializes an OAuth-backed remote MCP session and returns its tool schemas. |
latch_mcp_call |
Invokes a discovered remote MCP tool while retaining the session and keeping OAuth inside Latch. |
latch_capabilities is the one that changes agent behaviour. An agent that can ask
what it is permitted to do will stop attempting things it cannot do, which turns a
wall of denials into a short list of options.
Without MCP
Any HTTP client works. Set the base URL to https://onlatch.com/proxy and the bearer
token to the lat_… token. Most SDKs let you override the base URL, which is the
entire integration:
client = OpenAI(
api_key="lat_9f3c…", # the latch, not your OpenAI key
base_url="https://onlatch.com/proxy",
)The well-behaved agent never knows Latch is there. It just has less power than it thinks.
Add hardware identity to inference traffic
For an Anthropic Latch with an Identity filter, the Connect page also shows the
two environment variables needed for inference identity. Install Latch for Mac
v1.3.0 or newer first. It runs a signer on 127.0.0.1:8766; Claude sends the
existing lat_... token on each loopback request, and the service neither stores
the token nor passes it to another local process. The signer assigns the logical
Claude Code agent a stable K_a, signs the exact request, and forwards it to
onlatch.com/proxy. Users continue to start Claude with plain claude.
Native, npm/Node, and modified clients have the same default behavior. Executable/CDHash checks remain available only through the optional strict binary-verification mode. The one-off CLI launcher continues to support OpenAI-compatible clients when their command is supplied explicitly.
Projects connected before v1.3.0 require a one-time refresh. Return to the
Latch's Connect → Claude Code page, rerun step 1's curl command to replace
the locally saved /latch-setup-… file, and then run the refreshed slash
command from step 2. Fully quit Claude Code and VS Code before reopening them so
their existing MCP processes stop. The refreshed configuration uses
remote-mcp-2 and the stable claude-code subject; merely rerunning an older
saved slash command can reinstall remote-mcp-1 and retain the legacy
executable-based K_a instead of sharing the inference signer's identity.
This is additive: MCP configuration does not change, and users can continue to use the direct base-URL/token flow above when hardware identity is not required.
Related
- Proxy API: the full contract.
- The pipeline: what the agent's requests are evaluated against.