Connect Sourcegraph to Claude
Give Claude governed access to Sourcegraph through Runlayer, with identity, policy, scanning, and audit on every call.
Set up
Add Sourcegraph to Claude in four steps
Admins approve the server once. Everyone in Claude then gets it through the same governed connection.
- 1
Approve Sourcegraph in Runlayer
Add Sourcegraph from the catalog and choose who can use it and which of its tools are allowed.
- 2
Connect Claude to Runlayer
Admins install the Runlayer Plugin for everyone from Claude’s admin plugin settings, or users add it themselves. With direct tool listing turned on for your deployment, Sourcegraph’s tools show up in Claude’s own tool list.
Claude setup guide - 3
Sign in to Sourcegraph
Each user signs in to Sourcegraph with OAuth. Runlayer keeps the tokens out of the client’s configuration.
- 4
Use Sourcegraph from Claude
Sourcegraph’s approved tools appear in Claude. Every call is checked against policy and recorded.
Tools
Sourcegraph tools in Claude
Sourcegraph reports its tools when your organization connects it, so the list depends on your deployment.
Control
What Runlayer can see and stop in Claude
Some controls apply to every client at the gateway. Others depend on what Claude exposes to Runlayer.
Gateway policy (available)
Every Sourcegraph call from Claude passes through Runlayer, where your policies can allow, alert, mask, or block it.
Audit trail (available)
Each call is recorded with the user, Claude as the client, the tool, and the policy and scanner decisions.
Anthropic inference hooks (available)
With Claude Enterprise, Anthropic sends each governed Claude prompt to Runlayer before inference, with nothing installed on devices (Anthropic beta).
Compliance API import (available)
Anthropic Compliance API imports Claude web and desktop chats for review.
Managed browser extension (available)
Managed Chrome, Edge, and Firefox deployments can monitor supported web-chat activity and apply prompt-time controls where the surface permits it.
Stop a call before it runs (available)
Runlayer can reject a Claude prompt before Anthropic runs it, which stops the whole request rather than one tool call.
Block output after it runs (available)
Because Runlayer checks the full prompt, tool results in it can be blocked before the model sees them.
Redact output in place (not available)
Anthropic Inference hooks cannot return a rewritten prompt, so Runlayer cannot mask content in Claude. The gateway still applies policy to the call.
Frequently asked questions
Connect more