Connect your agent

Start with Quick start if you want copy-paste prompts. This page is the creator how-to for MCP: connect Claude, Cursor, Copilot, or another client, then either build or revise a draft in chat or run Deployed Listed decision trees against evidence. For the public wire-level tool catalog, use /mcp.

Need help? MCP Tools for authoring and tool questions; MCP Clients for connector, OAuth, or client setup.

Build or revise a draft in chat

Your agent constructs the dt_graph_json using SMEme’s design guidance. SMEme validates and saves the agent-constructed graph — the MCP authoring tools do not generate the graph, and they never automatically Deploy or change Listed / Hidden.

Call smeme_authoring_design_guidance before drafting.
You and your agent agree the questions, options, branches, and conclusions. The agent then constructs the dt_graph_json from that design using SMEme’s guidance.
Call smeme_authoring_validate_graph repeatedly, fixing findings, until draft_ready is true.
For a new tree, call smeme_authoring_create_draft (strict: requires draft_ready). To revise an existing draft, call smeme_authoring_get_draft, then smeme_authoring_update_draft with expected_graph_hash. Update may save an intentional intermediate graph; on graph_conflict, re-fetch and retry.
Open the editor, review the saved draft, then Deploy and choose Listed or Hidden. Review, Deploy, and Listed/Hidden are editor actions — separate from MCP authoring saves.

Authoring and tool questions: Ask in MCP Tools. Connector, OAuth, or client setup: Ask in MCP Clients.

Before you run a deployed tree

OAuth has succeeded. If the first MCP call returns auth_error, complete the web sign-in or legal consent the message names, then reconnect once.
You do not need a tree of your own first. The first empty smeme_reasoning_list creates a per-user sample (Deployed + Listed). You can also click Load sample on the dashboard. To run a tree you authored, Deploy it and set Listed. See Deploy, list, & delete.

Setup

Add the MCP connector

In your AI client, add a custom MCP connector (remote HTTP) with the values below, then complete OAuth when prompted.

MCP connector values

Copy these when adding a remote connector in your AI client.

MCP endpoint URL

https://app.smeme.ai/api/v1/mcp

OAuth Client ID

When your client asks for a Client ID (leave secret blank — PKCE public client):

NRdsdBvrio0DW9yo

Step-by-step by client

Claude (Chat, Desktop, and related surfaces)

Open Customize → Connectors at claude.ai/customize/connectors (same list in Claude Chat and Claude Desktop).
Click + → Add custom connector.
Paste the MCP endpoint URL from above.
Open Advanced settings → enter the OAuth Client ID from above. Leave client secret blank (PKCE public client).
Click Add, then complete OAuth. Sign in with the same email you use on SMEme.
In a conversation, click + → Connectors and turn SMEme on for that chat.

Team / Enterprise: an Owner adds the connector under Organization settings → Connectors first; you then click Connect. Anthropic help.

ChatGPT

Requires ChatGPT Pro, Team, Enterprise, or Edu with Developer Mode. Your agent bootstraps via guidance_get after OAuth. Initial setup is browser-only: turn on Developer Mode and create the connector at chatgpt.com — not in the ChatGPT desktop app. After OAuth succeeds, the connector is on your account and SMEme will be available across all platforms and devices.

In your browser, open chatgpt.com → Settings → Apps & Connectors.
Scroll to Advanced settings at the bottom of that page → turn Developer mode on and accept the warning.
Click Create (appears after developer mode is on).
Enter a name (e.g. SMEme), paste the MCP server URL from above, and set Authentication to OAuth.
In the create dialog, open Advanced settings → paste the OAuth Client ID from above. Leave client secret blank (PKCE public client). SMEme requires this static Client ID — it is not shown on the main form until you expand Advanced settings here.
Click Create (or Scan tools if shown) → complete OAuth and sign in with the same account you use on SMEme.
Start a new chat, click + in the composer → enable your SMEme connector for that conversation.

Two “Advanced settings” panels: page-level (step 2) enables Developer mode; dialog-level (step 5) is where you paste the OAuth Client ID. If OAuth fails, confirm both are set and retry. Workspace admins may need to allow custom apps. OpenAI connect guide.

Other clients (Cursor, Microsoft Copilot, etc.): add a remote MCP server with the same URL and OAuth Client ID in that client’s MCP settings, then complete OAuth.

Let your agent bootstrap

After OAuth, your agent should call tools in this order (most capable agents do this automatically):

smeme_reasoning_capabilities — read guidance.content_digest and the tool list.
If there is no cached guidance or the digest changed → smeme_reasoning_guidance_get — cache content_markdown (full calling contract).
Guided (default): smeme_reasoning_list (first empty list may create the sample; pass list id as decision_tree_id) → smeme_reasoning_evaluate → loop smeme_reasoning_evaluate_continue until harness_next is not continue_evaluate (a report).
Bulk/audit: smeme_reasoning_template_get → smeme_reasoning_validate_answers → smeme_reasoning_evaluate_answers.

Public tool reference: /mcp.

MCP tools — run deployed trees

Your agent discovers these at connect time. Guidance tools (quota 0) bootstrap the session; the rest follow the sequence in step 2. Authoring tools for building drafts are listed under Build or revise a draft in chat.

Tool	What it does	Quota
smeme_reasoning_capabilities	Session bootstrap: service info, MCP surface version, authoritative tool list, and `guidance.content_digest`.	0
smeme_reasoning_guidance_check	Cheap guidance digest probe — use only when skipping capabilities.	0
smeme_reasoning_guidance_get	Full platform calling contract as markdown — cache when digest matches.	0
smeme_reasoning_list	Returns your Deployed + Listed trees. Rows use `id` (pass later as `decision_tree_id`). Sample rows also include `sample_key: smeme_sample_v1`. The first empty list may create that sample (one tree slot; not read-only). Persistent empty is recovery: click Load sample or Deploy + Listed.	0
smeme_reasoning_list_conclusions	Catalog of possible outcomes — used before `how_to_reach`.	0
smeme_reasoning_template_check	Lightweight worksheet drift check — does not return question bodies.	0
smeme_reasoning_template_get	Authoritative per-decision-tree worksheet — question IDs, labels, and valid option strings.	0
smeme_reasoning_validate_answers	Validates provenance envelope shape, refs, and consistency before bulk Apply.	1
smeme_reasoning_evaluate	Guided case gather start — returns a blind task (or a terminal `report`).	1
smeme_reasoning_evaluate_continue	Guided case gather continue — admit one answer; next blind task or terminal `report` (may include `stop_reason`; branch on `report.result_kind`).	1
smeme_reasoning_evaluate_answers	Bulk Apply on a worksheet envelope; returns `report` with `result_kind`.	1
smeme_reasoning_what_if	Scenario comparison: baseline vs override with structured delta.	2
smeme_reasoning_how_to_reach	Minimal answer edits to reach a target conclusion.	2.5

Quota weights count against your monthly MCP tool call allowance. See Plans & usage.

Invoking a decision tree

Once connected, prompt your agent in plain language. Mention the decision tree name or describe the task:

"Run the foreign national property tax decision tree on the Chen matter using the files in this project."

"Evaluate whether this lease qualifies as a triple-net lease."

"Use SMEme to determine eligibility and then draft the client letter."

Your agent should: bootstrap guidance → list decision trees (the first empty list may seed the sample) → guided evaluate / evaluate_continue until a report (or, for bulk/audit: worksheet → validate → evaluate_answers). To get a first report without an LLM, see Core examples/smeme_apply_sample.py.

Tips

Try the sample first. Click Load sample on the dashboard, or let the first empty MCP list create it. It is a sanitized walkthrough owned by you — not legal, tax, or other expert advice — and it uses one decision-tree slot.
Connector first. URL + OAuth Client ID + OAuth connects SMEme. After OAuth, your agent loads the calling contract via smeme_reasoning_guidance_get — there is no separate install package.
Name decision trees clearly. The name is how your agent finds the right tool — e.g. "Foreign national property tax analysis" rather than "DecisionTree 3".
Redeploy after edits. Changes saved in the editor don't take effect until you redeploy. If your decision tree shows Stale on the dashboard, agents run the old version.
Use Hidden for drafts. Keep a decision tree Hidden until it's fully tested. Listed decision trees are available to your AI immediately.
Ask follow-ups after evaluation. Stay in the same conversation for what-if, how-to-reach, or document drafts — scenario comparisons count as 2 quota units; repair plans as 2.5.

Example use cases

Legal matter analysis: Encode a checklist for a category of matter. Your agent gathers facts, calls SMEme, returns a determined outcome with the full reasoning chain.
Compliance & eligibility screening: Build decision trees for regulatory eligibility questions. Get a yes/no/conditional with the conditions that decided it.
Repeatable triage: Encode a triage procedure so every intake gets evaluated the same way.
Scenario analysis: Vary one input and re-run with smeme_reasoning_what_if or find minimal edits with smeme_reasoning_how_to_reach.