Tools & MCP
Tools & MCP is Exemplar’s out-of-the-box tool surface for connectors and integrations—optimized to run efficiently and effectively under governance, and usable the same way from agent frameworks or MCP clients (Cursor, Claude Code, Codex, and custom agents).
Connectors you enable publish as unified MCP tools on a single org endpoint. Credentials stay vaulted; calls share one identity, policy, and audit trail—so teams stop hand-wiring per-integration glue for every agent.
Open it in the console: console.exemplar.dev → Agent Harness → Tools & MCP. Integrations catalog: exemplar.dev/integrations .
Tools & MCP is the living gateway: browse platform tools, turn your own OpenAPI into MCP tools, discover connector specialist agents (A2A), and copy connect snippets for IDEs and frameworks. Manage connector credentials under Connectors; govern IDE coding-agent MCP allowlists with Relay MCP control.
Why it matters
Without a shared tool plane, every agent team rebuilds bridges to Jira, PagerDuty, Datadog, cloud, and internal APIs—different auth, no shared audit, and token-heavy ad hoc calls. Tools & MCP standardizes that surface:
- One endpoint —
{agent_base}/mcpexposes a flat set of unified tools (memory_tool,skills_tool,pagerduty_tool, …). - Same call shape — every unified tool uses
{ action, params }, so clients and frameworks share one mental model. - Governed by design — org API keys, connector credentials in the vault, and the same Harness / Relay policy layers as the rest of the platform.
- Efficient for agents — connectors publish tools automatically; agents get live actions without per-vendor SDKs in every runtime.
How it works
Connectors and platform capabilities flow into one MCP core; frameworks and IDE clients consume the same governed catalog—optionally with A2A specialists for delegated turns.
| Stage | What happens |
|---|---|
| Connect | Enable providers; credentials stay vaulted |
| Publish | Unified *_tool actions (+ optional OpenAPI tools) appear on /mcp |
| Consume | Point MCP clients or harness.tools() at one endpoint |
| Delegate | Orchestrators can hand a whole turn to a connector A2A agent |
| Govern | Same org key, policy, and audit — Relay still allowlists IDE MCP calls |
Two protocols
| Protocol | Role |
|---|---|
| MCP | Direct tool access for Cursor, Claude, LangChain, Agno, Google ADK, and custom MCP adapters |
| A2A (Agent-to-Agent) | Delegate a whole turn to a connector specialist (Jira, PagerDuty, …) via JSON-RPC + agent card |
Use MCP when the agent should call discrete actions. Use A2A when an orchestrator should hand off reasoning about one integration.
Topics
Related
- Harness — gateway, prompts, skills, memory, and guardrails that sit behind the same tools
- Context management — connectors and per-turn context assembly
- SDK —
harness.tools()and framework MCP adapters - Automation Studio — graph workflows that invoke connectors and agents
- Relay MCP control — allow/deny which MCP tools IDE coding agents may call
- Account Settings — Tokens — API keys for MCP clients