Product
Connect. Compose. Mint. Govern.
The complete tour of the Elaichi MCP control plane — from a 200+ connector catalog to permission-clamped endpoints your whole company can trust.
Connect
A catalog of 200+ connectors, ready on day one
Connect the SaaS tools your teams already use — owned at user, team, or org scope — through a hosted connect flow that vaults every credential.
Instant tools
Every connector ships with documentation, and every documented method is automatically an LLM-callable MCP tool. No tool authoring to get started.
Hosted connect flow
OAuth dances, token refresh, and per-account secrets are handled by a dedicated credential vault. Connection status (active / needs reauth) surfaces in the app, with one-click reconnect.
Custom connectors
Author a connector from a JSON config — base URL, auth, resources, methods — for any internal or niche API. Validation returns precise, path-qualified errors.
Fork & pull from upstream
Fork any public connector including its documentation. Forks track lineage, so a review surface shows new tools, safe updates, and conflicts — and you apply changes selectively.
Compose
Toolboxes: the unit of sharing and control
A toolbox is a curated set of tools across connections — renamed, re-described, clamped, and shared with exactly the right people.
Delegated connections
A shared toolbox can expose the sharer’s connection. Recipients call third-party APIs through it without ever seeing credentials — HR shares exactly the ethics folder, nothing else.
Frozen parameters
Fix tool arguments to values stripped from the advertised schema and force-merged at execution. Precedence is defaults < model args < frozen params — the model can’t see or override them.
Per-tool customization
Rename and re-describe what the LLM sees, disable entries, override input schemas wholesale, and set argument defaults — all per toolbox entry.
Synthetic tools
Define a composite tool as a DAG of steps across connections with templated arguments. Independent steps run in parallel, server-side, through the same restriction and audit pipeline.
Dynamic toolboxes
Every user gets an auto-generated global toolbox plus one per connection — useful from the second a connection goes active.
Mint
A personal MCP endpoint for every user
Mint an endpoint on any toolbox and point Claude Desktop, Cursor, or Dust at it. Standard MCP JSON-RPC over Streamable HTTP.
Token rotation
A lost URL isn’t a dead end. Rotate re-mints the token for the same server — old token revoked first — so clients are re-pointed, not rebuilt.
Context-blowup protection
Toolboxes exposing more than ~30 tools automatically advertise just search_tools(query) and execute_tool(name, args), keeping large toolboxes out of the model’s context window. Configurable per server.
Live resolution
Every tools/call resolves the toolbox at execution: overrides applied, frozen params injected, restrictions clamped, then the proxy call runs with vault-resolved credentials.
Rate limits
120 requests/min per MCP token and 600/min per API token, with proper 429 + Retry-After responses.
{
"mcpServers": {
"elaichi-hr-toolbox": {
"url": "https://api.elaichi.ai/mcp/elch_••••••••••••"
}
}
}
// One personal endpoint per user — clamped by roles, restrictions & frozen params.
Govern
Three layers of control, one audit trail
Who can act, what they can reach, and a record of everything — enforced server-side where the AI can’t argue with it.
RBAC
Five seeded system roles plus custom roles built from a fixed permission catalog. Role edits apply live to everyone holding them. Per-resource ACLs add view/use/edit sharing grants.
Restrictions
Allow/blocklists over connectors and individual tools, targeted at the org, a role, or a user — precedence: user override > role > org default. Enforced at connect, advertise, and execute time.
Audit log
Append-only, per-org, covering memberships, roles, connections, shares, MCP lifecycle — and every tool call with tool, connection, status, and duration. Forward to your Datadog.
Offboarding preflight
Removing a member forces resolution of personal connections that shared toolboxes depend on — transfer or delete, never a silent break.
AI operator
An agent that drives the product itself
A chat agent lives inside the console and can actually do the work — create toolboxes, set entries, share, connect — through an allowlisted operation catalog gated by RBAC and per-action approvals (allow once, always allow, deny). It knows what page you’re on, can navigate the UI, and previews unsaved drafts live before you hit save.
- ✓ Bring your own LLM keys: Anthropic, OpenRouter, Fireworks, or any OpenAI-compatible gateway
- ✓ Every write goes through the same RBAC and audit pipeline as the UI
- ✓ WebMCP: browser agents can drive Elaichi through the same allowlisted operations
- ✓ Voice, @mentions, slash commands, and context compaction built in
Desktop companion
From toolbox to Claude in one click
The menubar app lists your toolboxes and MCP servers, and patches claude_desktop_config.json and Cursor’s mcp.json directly — add or remove server entries without anyone hand-editing JSON. Tokens are encrypted with the macOS Keychain.
- ✓ Claude Desktop and Cursor supported first; the config writer generalizes to more clients
- ✓ Mint and revoke MCP servers from the tray
- ✓ Built for non-technical teammates
Give every teammate a governed MCP server
Start a 14-day Gold trial — connect a product, curate a toolbox, and paste an endpoint into Claude or Cursor in minutes.