Not another walled agent platform. Switchboard is the open-protocol gateway under your agents: Cursor, Claude Code, Codex, or any MCP client on top; GitHub, Datadog, Slack, private data, and 30+ more underneath. Policies, audit, tunnel, and up to 90% token savings in the middle.
Free forever for solo use · Source-available · Open MCP protocol
Any MCP client connects through Switchboard to your tools. Clients include Cursor, Claude Code, Codex, and Windsurf. Switchboard provides policies, token savings, a private-data tunnel, and an audit log. Tools include GitHub, Datadog, Slack, Postgres, and bring-your-own integrations.
{
"mcpServers": {
"switchboard": {
"url": "https://your-org.switchboard-mcp.com/mcp"
}
}
}Platform independent
Switchboard sits under your agents and above your tools. Swap clients, models, or how you run the gateway - your policies and integrations stay put.
Cursor today, Claude Code tomorrow, a custom agent next quarter. One MCP URL works everywhere the protocol does - no proprietary SDK tax.
We are the tool and control plane, not the model. Use Claude, GPT, Gemini, open weights, or whatever your client already routes to.
Thirty-six plus built-ins, OpenAPI or GraphQL imports, and Wasm plugins. Not a single vendor's closed catalog - your stack, on your terms.
Run managed Switchboard Hosted, or self-host the source-available core. Deploy it your way - inspect the code that touches credentials either path.
Integrations
Connect your entire stack. Every integration is available through a single MCP endpoint - bring your own when you need more.
What you get
Lock-in-free by design - with the controls teams actually need once agents touch real systems.
One MCP endpoint replaces dozens of separate tool configs. Your AI client gets GitHub, Datadog, Slack, Linear, AWS, Google Workspace, and 30+ more through a single connection.
Compaction strips redundant fields, columnarizes tabular data, and renders markdown before responses reach the LLM. A built-in dashboard tracks tokens and dollars saved in real time.
Organizations, departments, SSO, and fine-grained access. ABAC policies and glob tool-filtering decide exactly which member's agent can reach which system - and the audit log proves it.
Query Postgres and Kubernetes behind your firewall through a customer-run, outbound-only agent. No inbound ports, no credentials stored in the cloud, read-only enforced on your side.
Import any OpenAPI or GraphQL spec, or ship custom Wasm plugins. They get the same compaction, policy controls, and audit logging as built-in ones.
Inspect every line that touches credentials. Self-host the core under Elastic License 2.0, or use managed Hosted - same open MCP surface either way.
The shift
The difference is obvious the moment you open your config.
Context efficiency
Most gateways dump every tool definition into the prompt. Switchboard exposes just two meta-tools - so adding your 40th integration costs nothing on unrelated calls.
Your agent finds the capability it needs in plain language. Only matching tools come back - not the entire catalog.
Run any tool by name, or chain several across integrations server-side. Intermediate results never round-trip through the model.
Set shared context once and inject it automatically. Pin results by handle so the model reuses data instead of re-fetching it.
Who it's for
One gateway, three very different problems solved.
Roll out MCP to the whole org from one place. Add or remove integrations centrally, push config without touching a single developer's machine, and onboard a new hire with zero credential sharing.
Credentials encrypted at rest, ABAC policies deciding who reaches what, and an audit log that proves it. Reach private data through an outbound-only tunnel - no inbound ports, no keys in the cloud.
Skip the per-tool setup dance. One endpoint in Cursor or Claude Code and your agent can reach everything - with ~90% fewer tokens per call. Free forever for solo use.
Compare
DIY MCP works until you have a team. Vendor agent hubs work until you want out. Switchboard is the portable middle.
| Raw MCP servers | Vendor agent hubs | Switchboard | |
|---|---|---|---|
| Client choice | Whatever you wire | Often one suite | Any MCP client |
| Model choice | Your client's | Often bundled | Yours - we never own the model layer |
| Setup per developer | One config entry per tool | Per-tool connectors | One endpoint URL |
| Credential storage | Plaintext local files | Vendor cloud | AES-256-GCM, self-hostable |
| Context / token cost | Full tool catalog in prompt | Varies | search + execute meta-tools |
| Access control | None | Basic roles | ABAC + glob tool filtering |
| Audit log | None | Some | Every mutation & tool call |
| Data behind your firewall | Manual | Rarely | Outbound-only read-only tunnel |
| Exit path | Your configs | Usually closed | Source-available core (ELv2) |
Architecture
Switchboard sits between your AI clients and your tools - auth, routing, and context optimization on an open protocol.
Need to reach data behind your firewall? A customer-run tunnel agent dials out to Switchboard, so agents can query private Postgres and Kubernetes with read-only enforcement on your side - no inbound ports, no credentials stored in the cloud.
Pricing
Free for individuals. Flat monthly team plans - no per-seat surprises. Most teams save more in tokens and consolidated tools than the plan costs.
Start free on Hosted, or self-host the source-available core. No credit card required.