Thinking Machines introduced Inkling as an open-weights, multimodal model family built for broad customization, including through Tinker. That makes it interesting for trading teams that want more control over the model layer—but control only helps when the model is given current market evidence rather than asked to improvise from memory.
For FX research, the useful division of work is straightforward: FXMacroData retrieves the current macro evidence; Inkling compares possibilities and turns that evidence into a brief a trader can inspect. Start with a release calendar, recent US inflation history, and USD/JPY context. Then require a result that keeps facts, scenarios, uncertainty, and review notes separate.
Fit
Use this for
Pre-release briefings, event monitoring, post-release reviews, scenario notes, and research assistants.
Inkling works best when
The surrounding host retrieves current FXMacroData evidence before the model is asked to interpret a setup.
Hand off to
A trader or analyst for the desk’s existing risk and decision process after the brief has been checked.
Why Inkling Fits FX Macro Research
Macro research is context-heavy. A desk needs the next event, prior prints, the current pair context, and a clean explanation of what would confirm or invalidate a view. Inkling is designed as a broad base model that can be customized, so it is a useful candidate for the reasoning side of that process. Its open-weights release and Tinker fine-tuning path make it particularly relevant when a team wants to own its prompts, evaluation set, and model adaptations.
The model is not the live data source. If the question is “what is the next inflation event?” or “what did the last release show?”, the answer should come from FXMacroData tools. If the question is “what are the conditional implications for USD/JPY?”, Inkling can synthesize the evidence into scenarios. Keeping those roles separate is what makes the output useful in a trading workflow.
The Evidence-to-Scenario Workflow
1. Ask
Frame the research problem around an event, pair, or policy question—not a generic request for a market view.
2. Retrieve
Collect current calendar rows, announcement history, pair context, and freshness or provenance details from FXMacroData.
3. Reason
Ask Inkling to distinguish observed facts from scenarios, interpretation, and uncertainty.
4. Review
Validate the brief, inspect its evidence trail, and hand it to the desk’s existing risk and decision process.
Trading-desk scenario
Before a US inflation release, a strategist asks for USD/JPY event risk. The host retrieves the release timing, recent inflation rows, and pair context. Inkling then writes three conditional paths: stronger-than-expected inflation, weaker-than-expected inflation, and a no-trade case where the evidence is mixed or incomplete.
MCP First, REST When the Application Needs Control
Inkling is a model, not an MCP client by itself. The MCP connection belongs in the agent host or application around the model. When that host supports remote MCP, use FXMacroData MCP as the first AI-facing data path. It gives the host a discoverable tool surface for exact macro rows, calendars, FX data, and research workflows.
| Route | Use it when | Why it matters |
|---|---|---|
| FXMacroData MCP | The host around Inkling can use a remote MCP server. | It is the first AI-facing path: the model can work with a focused, discoverable macro tool surface. |
| FXMacroData REST | Your application owns the credentials, data checks, cache, or non-MCP integration. | It is the controlled fallback for retrieving a compact evidence package before the model receives it. |
| Dashboard handoff | A person needs to inspect the context visually. | Send the desk to a pair dashboard, the release calendar, or Federal Reserve releases before deciding what to do next. |
Step 1: Connect an MCP-Capable Host
What to do: point the agent host around Inkling at the hosted FXMacroData MCP endpoint. The exact configuration shape depends on the host, but the endpoint is the same:
https://mcp.fxmacrodata.comA host that uses HTTP MCP entries shaped like this can start with the following configuration:
{
"servers": {
"fxmacrodata": {
"type": "http",
"url": "https://mcp.fxmacrodata.com"
}
}
}Give the model a compact starting tool set: data discovery, indicator queries, release calendars, and FX context. The FXMacroData MCP documentation explains how to check freshness, provenance, coverage, and point-in-time fields before a result is used in a research note.
Step 2: Add a Focused REST Fallback
What to do: when the host cannot call MCP directly, retrieve a narrow evidence package in your application and pass the resulting facts to Inkling. This keeps the current-data step explicit and avoids treating free-form web access as a research source.
import os
import requests
API_ROOT = "https://api.fxmacrodata.com/v1"
def fxmd_get(path: str, **params) -> dict:
params["api_key"] = os.environ["FXMD_API_KEY"]
response = requests.get(f"{API_ROOT}{path}", params=params, timeout=20)
response.raise_for_status()
return response.json()Use focused production requests with query-parameter authentication:
curl "https://api.fxmacrodata.com/v1/calendar/usd?api_key=YOUR_API_KEY"
curl "https://api.fxmacrodata.com/v1/announcements/usd/inflation?api_key=YOUR_API_KEY"
curl "https://api.fxmacrodata.com/v1/forex/usd/jpy?api_key=YOUR_API_KEY"Then build a small context object for one question rather than exposing a broad, unbounded data surface:
def usd_jpy_inflation_context() -> dict:
return {
"usd_calendar": fxmd_get("/calendar/usd"),
"usd_inflation": fxmd_get("/announcements/usd/inflation", limit=6),
"usd_jpy": fxmd_get("/forex/usd/jpy", limit=30),
}Step 3: Require a Reviewable Research Brief
Give Inkling a clear output contract. The goal is not an all-knowing market verdict; it is a compact briefing that lets a trader see what is known, what is conditional, and what still needs checking.
You are an FX macro research assistant.
Use the supplied FXMacroData evidence before interpreting current conditions.
Separate observations, conditional scenarios, missing data, and review notes.
Preserve event timing and source context.
Do not turn the brief into an execution instruction.A useful user prompt is equally specific:
Build a pre-release USD/JPY briefing for the next US inflation event.
Use the attached FXMacroData calendar, inflation history, and pair context.
Return stronger-inflation, weaker-inflation, and no-trade scenarios.
Show the evidence used and anything the desk should check next.Illustrative output shape
{
"observations": ["Current macro rows retrieved from FXMacroData"],
"scenarios": ["Stronger inflation", "Weaker inflation", "No-trade"],
"missing_data": [],
"review_notes": ["Confirm the release result before acting"]
}Step 4: Fine-Tune Only After You Have an Evaluation
Thinking Machines makes Inkling available for fine-tuning through Tinker. That can become valuable when a team has a repeatable research format it wants the model to follow—for example, consistently separating observations from scenarios or producing the same review fields for every event note.
Start with prompting and tool retrieval first. Save a small set of real research tasks, define what a complete briefing looks like, and test whether the model retrieves and structures evidence correctly. Fine-tuning is more useful once you can measure the behavior you want to improve than when it is used as a substitute for a workflow design.
Workflow Controls for Trading Research
| Control | What it protects | Practical rule |
|---|---|---|
| Evidence first | Current facts stay separate from model memory. | Retrieve calendar rows, data, and quality fields before asking for interpretation. |
| Scoped tools | The model stays focused on the macro research question. | Start with releases, indicators, and FX context before expanding the tool set. |
| Conditional scenarios | The brief does not collapse uncertainty into one directional view. | Always require bullish, bearish, and wait-or-no-trade cases when the data supports them. |
| Human handoff | Research stays distinct from the desk’s trade-decision process. | Route the completed brief to a trader or analyst for review. |
What to Measure After Launch
Evaluate this workflow as a research process rather than a demo. A strong result is one that retrieves the right evidence, identifies what it does not know, and saves a trader time without blurring the handoff to the desk.
| Measure | Good sign | Warning sign |
|---|---|---|
| Evidence use | Each brief uses current FXMacroData rows and preserves the event context. | The model writes an answer without fresh data or omits what it retrieved. |
| Scenario quality | The result compares conditions that would strengthen, weaken, or defer a view. | Every question receives one confident directional answer. |
| Review speed | A trader can see facts, scenarios, and open questions in one short brief. | The output is long, unsourced, or difficult to challenge. |
Common Questions
Can Inkling be used for FX trading research?
Yes. Inkling can act as the reasoning layer in a workflow that retrieves current macro evidence, compares scenarios, and produces a structured brief for review.
Should an Inkling workflow use MCP or REST with FXMacroData?
Use FXMacroData MCP first when the host around Inkling supports remote MCP. Use REST behind your application when it needs direct control over credentials, validation, caching, or a non-MCP integration path.
Does this let Inkling place trades?
No. This workflow is for research and scenario analysis. Keep trade decisions and execution within the desk’s separate review, risk, and order controls.
What is a useful first Inkling workflow for traders?
Start with a pre-release briefing: retrieve the next event, recent announcement history, and pair context, then ask Inkling for conditional scenarios, missing data, and follow-up checks.
Related FXMacroData guides
Sources
This guide is educational engineering content, not investment advice or a recommendation to trade any instrument.
Bottom Line
Inkling is a strong fit for the reasoning side of an FX research workflow when the model is grounded in current evidence. The practical pattern is simple: use FXMacroData MCP first for AI-facing retrieval, keep REST as the controlled application fallback, ask Inkling for conditional research rather than an execution instruction, and preserve a clear handoff to the desk.
