Blazle Conduit · Agents and tools: MCP gateway and approvals
Conduit. Every tool call, on your terms.
Conduit sits in front of your MCP servers. Agents reach tools only through it, and risky actions wait for a person.
The problem
- Agents can call any tool an MCP server exposes
- Tool results can carry instructions that steer an agent
- Destructive actions run before anyone can say no
How it works
Conduit. Every tool call, on your terms.
- 01
Register your MCP servers
Agents connect to the gateway with their own MCP token.
- 02
Scan each call
Tool names, arguments and results go through the same interceptor.
- 03
Approve what matters
State-changing calls wait for a person; destructive ones need a second approver.
| Event | Source | Result |
|---|---|---|
| Conduitgithub.delete_repositoryrelease-agent | release-agent | Waiting for approval |
| Conduitshell.exec · rm -rf /ops-agent | ops-agent | Blocked |
| Conduitjira.create_issuetriage-agent | triage-agent | Allowed |
Controls
Per-tool allow, block or approve
Switch off any tool with one click.
Four-eyes approvals
A second person for destructive actions.
Output guard
Results are checked before they reach the agent; secrets are masked, instructions hidden in data are labelled untrusted.
Full audit
Proposal, decision and execution are all recorded.
Plans and price
- 1 MCP server on Hacker and Pro, 5 on Business, unlimited on Enterprise.
- Extra servers: $49 per server per month ($490 per year), from Pro.
