Complete intelligence runtime for agents — not another search API.
Embed Research, Context, Memory, Control, and Pipeline once via MCP/API instead of rebuilding fetch, dedup, compression, budget guard, memory, and orchestration per agent. Seven smart_* primitives sit on one Streamable HTTP endpoint for Cursor, Claude Desktop, or your runtime. Copy MCP config by host, same as Dashboard Connect.
Seven algorithm primitives (MCP)
Mechanism layer for the intelligence runtime — same names and semantics as MCP docs. One URL; your host LLM keeps reasoning.
- smart_fetchread-only
Fetch URL as Markdown
Fetch a public HTTP(S) URL and return Markdown-friendly content.
When: User provides a link and needs the page body — not when text is already in chat.
Composes: smart_context_gate on long pages; smart_search when no URL yet.
- smart_searchread-only
Web search
Multi-engine web search with aggregated snippets.
When: Research without a definitive URL — then fetch the best hit.
Composes: smart_fetch on results; smart_pipe research template.
- smart_context_gateread-only
Compress context
Compress long context (LLMLingua) before the host LLM call.
When: Context exceeds model budget; optional purpose string filters paragraphs.
Composes: Usually follows smart_fetch or smart_search.
- smart_dedupread-only
Semantic deduplication
Semantic dedup across overlapping passages (SemHash).
When: Several chunks or search results overlap — not a single string.
Composes: After smart_search; before smart_context_gate.
- smart_budget_guardread-only
Budget check/count/record
Team token budget check, count, or usage record.
When: check/count before expensive tools; record after token work.
Composes: Before large fetch/compress jobs (API key scopes tenant).
- smart_memory
Team memory
Team-scoped memory — add, search, get, delete.
When: Persist facts across sessions — not world encyclopedia answers.
Composes: smart_search for grounding; smart_pipe remember template.
- smart_pipe
Pipeline orchestrator
Multi-step pipeline — templates research, read, remember, or custom steps.
When: Several tools in sequence — not a single direct tool call.
Composes: steps[].params with smart_* tool names per step.
One-click MCP config
Same host tabs as Dashboard → Connect — copy JSON, then paste your API key.
Streamable HTTP (MCP spec 2025-03-26): POST /api/mcp with Bearer API key and agent-platform header — same config as Connect.
Works with any MCP host that supports Streamable HTTP and Bearer auth (Cursor, Claude Desktop, Windsurf, Hermes, OpenClaw, and others). Endpoint: https://smartgate.network/api/mcp. Supported hosts
Config file: ~/.cursor/mcp.json (global) or .cursor/mcp.json (project)
{
"mcpServers": {
"smartgate": {
"type": "streamable-http",
"url": "https://smartgate.network/api/mcp",
"headers": {
"Authorization": "Bearer sk_live_REPLACE_WITH_YOUR_KEY",
"X-SmartGate-Agent-Platform": "cursor"
}
}
}
}- Official docs: Cursor MCP setup.
- Save as ~/.cursor/mcp.json (global) or .cursor/mcp.json (project). Merge smartgate into existing mcpServers if the file already exists.
- Env interpolation:
${env:SMARTGATE_API_KEY}in headers (Cursor resolves url and headers). - Restart Cursor or toggle the server under Settings → MCP after editing. SmartGate uses Streamable HTTP (stateless POST).
- Include X-SmartGate-Agent-Platform: cursor in headers (included in the snippet below) so Audit Logs shows the correct agent platform.
- Replace
sk_live_REPLACE_WITH_YOUR_KEYwith your real key — no curly braces around the token. - Use host IP instead of
localhostwhen the MCP client runs on a different machine. - See MCP endpoint guide for smart_* tools and troubleshooting.
- Endpoint: POST
/api/mcp(Streamable HTTP, stateless, no session management needed). - X-SmartGate-Agent-Platform in headers labels the agent in Audit Logs (Dashboard → Logs). Included in each platform snippet below.