The Model Context Protocol lets an AI application discover and call external tools through a standard interface. Nebula's MCP server gives compatible agents access to the same equity-first market-intelligence operations exposed by the public API, including asset resolution, sentiment, mindshare, social momentum, insights, and market context.
The reliable pattern
Connect the server, resolve the stock before requesting stock-specific data, make the time window explicit, retain tool evidence, and require the agent to separate observations from interpretation.
What MCP does
MCP is an open standard for connecting AI applications to data sources, tools, and workflows. An MCP server publishes discoverable tools with input schemas. A compatible host can present those tools to a model, execute authorised calls, and return the structured results to the conversation.
Nebula's MCP server is generated from the public API's OpenAPI contract. Tool names, descriptions, parameters, constraints, and response schemas come from that source of truth. The available market tools include list_assets, get_sentiment, get_mindshare,get_social_momentum, list_insights, and other supported operations.
1. Connect an MCP client
The server endpoint is https://nebula-api.hiddensystems.ai/mcp. OAuth-capable clients can discover Nebula's OAuth flow. Existing direct clients may authenticate with a Nebula API key as documented in the MCP server guide.
claude mcp add nebula \
--transport http \
https://nebula-api.hiddensystems.ai/mcpKeep credentials out of prompts, source files, browser bundles, screenshots, and conversation logs. Review the client's permission model before enabling tools in a shared workspace.
2. Resolve the subject first
For a stock workflow, call list_assets with the equity name or ticker, then pass both the returned asset_id and asset_class to asset-scoped tools. This avoids guessed identifiers and ticker ambiguity. A good agent instruction is: “Resolve NVIDIA as an equity before requesting any stock-specific metric.”
People, accounts, and topics are non-asset subjects and use list_subjects. Do not represent a private AI business as a public stock or merge a business, product, and executive into one subject. Correct identity is a prerequisite for correct research.
3. Build a structured research workflow
Ask the agent to perform a sequence rather than request a broad conclusion in one step. A useful stock brief can resolve the equity, retrieve sentiment and mindshare over a stated window, examine social momentum, retrieve relevant insights, and then summarise what changed with evidence and uncertainty.
Research recent changes in NVIDIA market conversation.
1. Resolve NVIDIA as an equity.
2. Request sentiment, mindshare, and social momentum for the same seven-day window.
3. Retrieve relevant recent intelligence.
4. State the data timestamps and sample context.
5. Separate tool observations from your inferences.
6. Identify conflicting evidence and what should be verified in primary sources.
7. Do not provide a buy or sell instruction.The prompt defines the research contract, while the tool schemas define valid inputs. Keep prompts focused on process and output quality rather than hard-coding endpoint details that can drift.
Useful tool combinations
| Research task | Tool sequence | Human check |
|---|---|---|
| Stock sentiment brief | list_assets → get_sentiment → get_mindshare | Underlying claims and primary evidence |
| Unusual activity review | list_assets → get_social_momentum → list_insights | Duplication, catalyst, and timestamp |
| Event preparation | list_assets → list_calendar_events → sentiment tools | Issuer materials and event timing |
| Source research | list_assets → get_notable_calls → list_insights | Attribution, incentives, and track record |
| Market context | sentiment tools → price and volatility tools | Window alignment and valuation context |
4. Preserve evidence and provenance
Store or display the tool name, input, response time, subject identity, and window behind an important conclusion. Require the agent to cite returned evidence where available and to label any statement that is an inference. If the tool returns thin or missing data, the answer should say so.
An agent may combine a true metric with an unsupported explanation. For example, rising mindshare is an observation; “institutional investors are accumulating” is a separate claim requiring evidence. A structured output with sections for observations, interpretation, verification needs, and uncertainty makes that boundary visible.
5. Add operational safeguards
- Use least-privilege credentials and the host's approval controls where appropriate.
- Set bounded task scope, lookback windows, and maximum tool calls.
- Handle authentication, entitlement, rate-limit, validation, and transient errors separately.
- Do not let the model silently substitute a different stock or subject after a failed resolution.
- Expose freshness and suppress confident summaries when required calls fail.
- Prevent untrusted retrieved text from overriding the agent's tool and safety instructions.
- Require human review for investment, compliance, or external-publication decisions.
6. Evaluate the agent
Create a small set of representative tasks using NVIDIA, Amazon, and less-discussed equities. Include ambiguous tickers, thin data, conflicting narratives, API errors, and events where conversation follows price. Score whether the agent resolved the correct identity, aligned windows, used the right tools, retained evidence, and calibrated uncertainty.
Evaluate process separately from prose. A fluent answer with the wrong asset is a failure. A concise answer that says evidence is insufficient may be correct. Repeat evaluations after tool-description, model, or prompt changes and preserve failing examples as regression cases.
Practical agent applications
- Morning watchlist briefs with changes in sentiment, mindshare, and narratives.
- Earnings preparation that identifies the current investor debate before reading filings.
- Event triage that turns an unusual-activity alert into a sourced research packet.
- Coverage monitoring that highlights which holdings require analyst attention.
- Research notebooks that retain tool observations alongside the analyst's thesis.
For REST implementation, read the stock sentiment API guide and stock alerts guide. For the full opportunity map, see financial market intelligence use cases.
Specifications and documentation
- Model Context Protocol introduction
- MCP tools specification
- Nebula MCP server documentation
- Nebula public API reference
Where financial attention becomes signal
Explore live market intelligence in Nebula or use the same structured signals in your own workflow.