MCP servers and connectors
Where agent tools come from — the built-in Nested MCP server, connectors, and pack MCP servers — and how the catalog and Code Mode save tokens.
Three sources of tools#
| Source | What it is | Who sets it up |
|---|---|---|
Built-in nested server | Platform tools: tasks and A2A, knowledge, formulas, file sharing, asking you a question, site sign-in | Always on |
| Connectors | Remote MCP servers your account signs in to — GitHub, Slack, Google, Atlassian, Sentry, and team custom connectors | You, on the Connectors page |
| Pack MCP servers | MCP servers declared in a pack | Pack authors; see Packs |
The connector catalog#
Attaching every MCP server to every session is expensive: with 11 servers attached, a session spent about 28,000 tokens on tool schemas before doing any work. Nested attaches only the built-in nested server (and any servers an operator marks as eager). Everything else sits behind four catalog tools:
| Tool | What it does |
|---|---|
nested_mcp_list | Lists the connectors available to this task, with a hint for when to use each |
nested_mcp_describe | Loads one connector’s tool schemas, as JSON or compact TypeScript declarations |
nested_mcp_call | Calls one tool on one connector |
nested_mcp_exec | Runs a short script that calls connectors (Code Mode) |
The agent still discovers every connector automatically; it just loads schemas only when it needs them. Operators can turn catalog mode off to attach every server.
Code Mode#
nested_mcp_exec runs a short JavaScript program in a sandbox inside the nested server. The script can call connectors in loops, paginate, filter, and aggregate, then return only the final result. Intermediate payloads never enter the model’s context.
// Connector and tool names come from nested_mcp_list and nested_mcp_describe.
const prs = mcp.call("github", "list_pull_requests", { repo: "acme/app", state: "open" });
return prs.filter((pr) => pr.draft === false).map((pr) => pr.title);The sandbox has only the mcp and console objects: no filesystem, network, timers, or module loading. Limits per run:
| Limit | Value |
|---|---|
| Timeout | 60 seconds default, 120 maximum |
| Connector calls | 25 |
| Result size | 64 KiB |
| One connector reply | 8 MiB |
| Memory growth | 256 MiB |
Parallel calls (mcp.callAll) | 4 default, 8 maximum |
Each run returns an audit trail of the connector calls it made.
Connectors#
A connector is a remote MCP server that your account connects to. A task only sees the connectors of the person who owns it, in that task’s team; delegated subtasks keep the same owner.
- Sign-in. OAuth connectors use the authorization code flow with PKCE. Atlassian and Sentry register their OAuth client automatically, so operators need no developer-console app for them. For GitHub, Slack, and Google, an operator enters the OAuth app once in the console.
- Scopes. Some connectors let you choose permissions. Required scopes are always included.
- Tokens. Connector tokens are encrypted at rest in the control plane. Agents never hold them: a loopback proxy in the agent pod adds the current token to each request and refreshes it when needed. A connector that needs a new sign-in shows Needs reconnection.
- Gateway connector. A Teleport connector lets agents list and call internal apps behind Teleport, over mutual TLS, without installing Teleport tools in the agent.
Team custom connectors#
Team owners and admins can add their own remote MCP servers for their team, with per-person OAuth or a shared header. Shared headers are limited to an allow list (authorization, x-api-key, and similar). Server URLs must use HTTPS and resolve to public addresses, and every connection re-checks the address, which blocks DNS-rebinding tricks against internal services.
The built-in tools#
| Group | Tools |
|---|---|
| Collaboration | nested_discovery_*, nested_a2a_* |
| Knowledge and memory | nested_knowledge_* |
| Formulas | formula_<name>, nested_formula_* |
| Files | nested_share_file, nested_s3_* |
| People | nested_request_user_input |
| Websites | site_login, site_challenge, browser tools |
| Connector catalog | nested_mcp_* |
site_login signs in to a website with a login you saved. The agent gets a status, never the password.
Questions
- Do agents see my connector tokens?
- No. Tokens stay encrypted in the control plane. A loopback proxy in the agent pod adds them to requests, so the coding agent never holds them.
- What is MCP Code Mode in Nested?
- A tool that lets the agent run a short sandboxed JavaScript program that calls connectors and returns only the final result, which saves tokens on multi-step work.
- Can I add my own MCP server?
- Yes. Team owners and admins can add a custom connector for their team, and pack authors can declare MCP servers in a pack.