Elis Token Saver Connector
Elis Token Saver is a remote MCP server that plans a goal into right-sized tasks so your AI host spends fewer and cheaper tokens, then meters the savings. Elis plans; your host executes — the connector never runs models on your behalf.
1) Add the connector in Claude.ai
- Open Settings > Connectors (or Customize > Connectors) and choose Add custom connector.
- Enter the connector URL:
https://api.tryelisai.com/api/mcp/claudeClaude discovers the server, and the first tool call opens the Elis sign-in and consent screen (OAuth 2.1). On the consent screen, you can select a billing scope: Personal (default, uses your wallet and personal allowance) or one of your organizations (uses the organization's wallet and pooled allowance; admins can view runs on the org observability page). Once you approve, the connector is linked to your Elis account with the selected scope.
2) Add the connector in Claude Code
claude mcp add --scope user --transport http elis-token-saver https://api.tryelisai.com/api/mcp/claudeAdding the server is only half of it — in production every Token Saver tool call is refused until the host holds an OAuth token, so an unauthorized server looks exactly like a missing one. Restart Claude Code, then authorize it:
/mcp -> elis-token-saver -> AuthenticateThe authorize step needs an interactive terminal — a headless run, a CI job, or a spawned subagent cannot open the picker or the browser. Do it once from a terminal and later sessions reuse the stored token. --scope user registers the server for every project; a server defined in a project's .mcp.json instead stays inert until that project approves it at first launch. Confirm with claude mcp list. Codex and ChatGPT use /api/mcp/openai; Gemini CLI uses /api/mcp/gemini; Antigravity uses /api/mcp/antigravity.
3) Install in Claude Desktop
Claude Desktop can install Elis Token Saver as a packaged extension — no Node.js install or manual configuration needed. The bundle ships its own MCP proxy and runs on Claude Desktop's built-in runtime.
- Download the extension bundle (.mcpb).
- Open Settings > Extensions and drag the downloaded file onto the install area ("Drag .MCPB or .DXT files here to install").
- Confirm the install and keep the default server URL unless you self-host Elis. The first tool call opens the same Elis sign-in and consent screen as the Claude.ai connector.
Prefer no download? Claude Desktop also supports the custom connector path from section 1 (Settings > Connectors > Add custom connector) with the same URL.
4) Authentication and account requirements
- Production requires OAuth sign-in with an Elis account before planning work. Server discovery stays open so your host can register the server first.
- The OAuth consent screen lets you select a billing scope: Personal (default) charges your personal wallet, drawing first on that wallet's own daily free grant; selecting an organization charges that org's wallet, drawing first on the org wallet's own pooled daily free grant, and org admins can see the runs.
- Billable saved-token overage draws from the same wallet — there is no separate paid Token Saver program gate. If the wallet has no free grant left and no way to be charged (no balance and no charge-ready subscription), the connector reports that billing setup is needed instead of disclosing tasks.
- Connector tokens are scoped: they work only for the Token Saver surface and are rejected by the general Elis API.
5) The tools
- start_elis_token_saver_plan — start a plan for a goal; returns a quote (estimated savings, task count) with no task content.
- submit_elis_task_plan — submit a host-authored task decomposition when no proven chain matched the goal.
- get_next_elis_task — commit the plan, report finished work, and receive every currently-unblocked task.
- finish_elis_plan — settle the plan, record usage measurement data and optional feedback.
- setup_elis_host_instructions — preview-only manifest of the local host files that make Elis the default: the always-on Token Saver rule, plus the
/eliscommand and the per-task executor agent that command spawns. It returns their contents for your host to write; it never writes anything itself.
6) Run every task with /elis
setup_elis_host_instructions hands your host three files: the always-on Token Saver rule, the elis executor agent that carries out a single assigned task on the model Elis sized it for, and the /elis command that drives the whole plan — start it, decompose only when Elis asks, run every ready task at its selected tier, report each result, and finish. Once your host has written them, drive your work with the command:
/elis add pagination to the orders tablesetup_elis_host_instructions installs /elis once. Use /elis on EVERY task — it starts the plan, sizes each task to nano, small, or frontier, and finishes the plan. Running it once and then working normally gives the savings back.
The command also works bare: typing /elis with no arguments uses your current request as the goal, so you can describe the work first and hand it to Elis after.
7) Data handling
Successfully finished plans may be stored privately in your organization or personal scope as proven chains so similar goals plan faster later. Opt any plan out with constraints.no_store. Host-reported usage is recorded as measurement and estimation data only — settlement always charges the reserved planning estimate. Details in the privacy policy.
Community sharing (opt-in): off by default. When you turn it on in Settings, proven chains and agents from your scope are redacted (PII, secrets, and org identifiers removed), scanned for safety, and shared only with other participants who have also opted in — and you can use theirs in return. Exclude a single plan from sharing with constraints.no_community, even while your scope's setting is on. See the privacy policy.
8) Cold starts
The production backend can scale to zero after a quiet period. If your MCP client times out on the first connection, the server is usually waking — retry in a few seconds, or open tryelisai.com/mcp to warm it first.