Fluint Loop MCP Server Documentation
What this is
Fluint's MCP server lets any AI assistant read your Loop briefs directly. Loop continuously generates briefs for your deals and the emails and meetings around them, and the server exposes them through five read-only tools.
Server URL | |
Transport | Streamable HTTP |
Auth | OAuth 2.1. You authorize with your Fluint account, and every call is scoped to your tenant and your permissions. |
Access | Read-only. All five tools are reads; none can write to a brief or to a connected CRM. |
Availability | Directory listings for Claude and ChatGPT are in review. Until they go live, connect with the server URL. |
Prerequisites
You need a Fluint account, a connected CRM with briefs already generated, and any MCP client.
Requirement | Detail |
Fluint account | Active account in the Fluint Platform |
Connected CRM | HubSpot, Salesforce or Dynamics. Only completed briefs are queryable; a brief still generating reads as not found. |
Client | Any MCP client: Claude, ChatGPT, Cursor, Claude Code and others |
Permissions | The server respects your Fluint role. You will not see briefs for records you cannot see in the product. |
Setup
Connect with the server URL today; it works in every client. Once the Claude and ChatGPT directory listings are live, connecting there becomes a one-click flow.
Claude (web, desktop, mobile)
Go to Settings, then Connectors.
Click Add custom connector.
Name it Fluint and enter the URL https://mcp.fluint.io/mcp.
Click Add, then Connect, and authorize with your Fluint account.
Open a new chat and confirm the Fluint tools appear in the tools menu.
Claude Code
claude mcp add --transport http fluint https://mcp.fluint.io/mcp
Then run /mcp and authenticate.
ChatGPT
Add a custom connector under Settings, then Connectors, using the same URL. Fluint implements the search / fetch retrieval pair that ChatGPT's deep research mode expects.
Cursor and other MCP clients
{ "mcpServers": { "fluint": { "url": "https://mcp.fluint.io/mcp" } } }Coming soon: connect from the directory
Once the listings are live in Claude and ChatGPT, go to Settings, then Connectors, search the directory for Fluint, and click Connect. Clients without a connector directory, such as Claude Code and Cursor, will keep using the server URL.
How Loop data is organized
Loop keeps one brief per entity: a deal, an email thread or a meeting. Each brief is built from one or more raw source records, such as a CRM deal or a meeting transcript, and links back to them.
Brief types
Brief type | Covers | Entity ref example |
| A CRM deal or opportunity: summary plus qualification predictors |
|
| A recorded meeting: problem statement, next steps, metrics, attendees |
|
| An email thread: summary plus correspondence status |
|
ID formats
Three kinds of ID appear across the tools. Always copy them from a previous tool result rather than building them by hand.
ID | Format | Example | Used by |
Entity ref |
|
|
|
Search ID |
|
|
|
Record ref |
|
|
|
Anatomy of a brief
Every brief has the same envelope; what varies by brief type is the sections object. Sections are keyed by name, and a brief can have fewer sections than its type supports, or none, when Loop hasn't produced them yet.
{ "entity_ref": "hubspot:deal:63351418389", "brief_type": "deal_brief", "title": "Bridger Photonics (2027 Renewal)", "abstract": "Bridger Photonics is a 2027 renewal deal valued at...", "sections": { "predictors": [ ... ] }, "source_records": [ { "ref": "hubspot_deal:63351418389", "observed_at": "2026-10-02T13:11:48.000Z" } ], "covers_through": "2026-10-02T13:11:48.000Z", "computed_at": "2026-10-02T13:43:21.000Z", "version_id": "abe825d3-1921-439a-952e-422e8041306c" }
Envelope fields
Field | Notes |
| The entity this brief describes. Pass it to |
|
|
| Record title, such as the deal or meeting name |
| The brief's narrative summary, written as prose |
| Structured content, keyed by section name. Shape depends on brief type. |
| The raw records the brief was built from. Pass each |
| Timestamp of the latest source data the brief reflects |
| When Loop generated this version of the brief |
| Unique ID for this version |
Deal brief sections
predictors is an array of qualification scores. Each entry carries a numeric value and the reasoning behind it.
{ "def_id": "default:economic_buyer_access", "label": "economic_buyer_access", "definition_version": 4, "calculation": { "value": 15, "reasoning": "Katie Stewart is a Manager of Operational Excellence, not the economic buyer..." } }Predictor ( | Measures |
| Strength and engagement of the named champion |
| Whether the economic buyer is identified and engaged |
| Whether pain is explicit and its business cost quantified |
| Whether metrics, targets or ROI are quantified |
| Whether the buyer's evaluation criteria are documented |
| Whether the approval steps and stakeholders are known |
| Whether legal, procurement and security steps are known |
| Competitive dynamics in play |
Meeting brief sections
Section | Shape |
| String: the customer's problem as discussed |
|
|
|
|
|
|
Email brief sections
Section | Shape |
| Thread classification, such as |
|
|
List sections carry an absence field that explains an empty list, for example not_discussed or nothing_outstanding. An empty string means the list is populated.
Tool reference
Five tools, ordered by how you'll reach for them. In Claude and other clients that call tools directly, start with find_briefs; search and fetch exist for ChatGPT-style retrieval.
find_briefs
Finds briefs by asking a question in plain language, such as "which renewals have unsigned order forms". It matches on both the words and the meaning of every brief, so you don't need to know entity types or IDs first.
Parameter | Type | Required | Default | Notes |
| string | Yes |
| The question in full. Full phrasing ranks better than keywords. |
| array | No | all types | Filter, e.g. |
| integer | No | 10 | Max 50 |
Returns hits[], each with entity_ref, brief_type, title, abstract, covers_through and score. Scores rank results within one response only; they are not confidence values and can't be compared across calls. There is no pagination.
get_brief
Reads one brief in full: the abstract, every section, and links to its source records.
Parameter | Type | Required | Notes |
| string | Yes | From |
| string | Yes |
|
| string | No | ISO-8601 timestamp. Returns the brief as it stood at that time. |
Returns the full envelope described in Anatomy of a brief. Only completed briefs are visible. A brief that doesn't exist, or is still generating, returns NOT_FOUND: no completed brief for this entity.
get_record
Fetches one raw source record, such as a CRM deal, an email or a meeting, by the ref listed in a brief's source_records. This is the only way to reach raw data, and it works by ID only; you can't search for records directly. Use it when a brief names something and you need the underlying field, like an amount, a close date or an email body.
Parameter | Type | Required | Notes |
| string | Yes | Exactly as |
Returns found, ref, entity_type, source (e.g. hubspot.Emails), observed_at, and fields, which holds the record's raw CRM properties.
search
Keyword search across briefs, returning IDs that fetch accepts. It exists for MCP clients that use a search-then-fetch retrieval pattern, notably ChatGPT's deep research mode.
Parameter | Type | Required | Default | Notes |
| string | Yes |
| Free text |
| integer | No | 10 | Max 50 |
Returns results[], each with id (format <brief_type>/<entity ref>), title, url and text, where text is the brief's abstract.
fetch
Retrieves one brief by an ID from search, as a document a research client can quote and cite.
Parameter | Type | Required | Notes |
| string | Yes | From |
Returns id, title, text (the abstract), url, and metadata with entity_ref, brief_type, covers_through, computed_at, version_id and source_records. fetch doesn't return sections; use get_brief when you need scores or next steps.
Typical flows
Most sessions go question, then brief, then evidence. Each step's output carries the ID the next step needs.
Claude and direct tool-calling clients
find_briefs(query: "which renewals have unsigned order forms", brief_types: ["deal_brief"]) → entity_ref: "hubspot:deal:63351418389" get_brief(brief_type: "deal_brief", entity_ref: "hubspot:deal:63351418389") → abstract, predictors, source_records[].ref: "hubspot_deal:63351418389" get_record(ref: "hubspot_deal:63351418389") → raw HubSpot deal fields
ChatGPT deep research
search(query: "Locatrix expansion", limit: 5) → id: "deal_brief/hubspot:deal:54956503342" fetch(id: "deal_brief/hubspot:deal:54956503342") → title, text, url, metadata
Use cases
The tools are strongest at answering questions about specific accounts and small sets of deals, with evidence. They don't enumerate every deal in a pipeline, so portfolio-wide audits work on a shortlist of up to 50 briefs at a time.
1. Pre-call prep
"I'm meeting Acme tomorrow. Where is the deal weakest, and what did we commit to last time?"
find_briefs surfaces the deal brief and recent meeting briefs together. get_brief on the deal returns predictors with reasoning, so the rep hears "economic buyer access is low because Katie is in operations, not budget" rather than a generic summary. The meeting brief's next_steps lists who owes what.
2. Follow-up audit
"What did we commit to on customer calls this month, and who owns each item?"
Meeting briefs carry next_steps with owner, due and committed. Checking them against later email briefs shows which commitments went quiet.
3. Renewal risk sweep
"Which renewals are stalled on signature or payment?"
A plain-language find_briefs question with brief_types: ["deal_brief"] finds deals by situation, even when the CRM has no field for it.
4. Qualification gaps on a shortlist
"Of my renewals this quarter, which score lowest on economic_buyer_access or decision_process_known?"
find_briefs builds the shortlist, then get_brief on each returns comparable predictor values. Some deal briefs have no predictors yet; treat those as unscored, not as low.
5. Evidence for a claim
"The brief says the order form went out on 26 August. Show me the email."
get_record on the brief's source_records returns the raw CRM record, so the assistant can quote the actual amount, date or message instead of the summary.
6. Coaching from reasoning
"Across these deals, what keeps coming up in the reasoning for low pain_validated scores?"
Each predictor's reasoning explains its score in plain language, which makes patterns visible, such as pain being described but never quantified.
7. CRM drift detection
"Does what the brief says about the FISCAL Tech renewal match HubSpot?"
Compare the brief against get_record, or pair the Fluint connector with your CRM connector in the same session.
