describe_tool
ChatGPTGet one tool's full detail by name: description, input_schema (JSON Schema of its params), and behavioural annotations. Pair with list_available_tools (call that first to find the name). An unknown name returns a note listing valid names — read it and retry, don't error.
find_company
ChatGPTFind a company by name or security identifier (LoanX id, CUSIP, ISIN, FIGI, Bloomberg) and get its Octus company id and canonical name. Call this first when a question names or identifies a company, then pass the returned company_id to the other tools. Deterministic three-way result: confident match → matched: true + company: {company_id, company_name, aliases}; good-but-uncertain → also candidates (other companies, each with aliases) so you can confirm the right one; no confident match → matched: false, company: null, and candidates to choose from.
get_company_details
ChatGPTGet an Octus company card: identity, GICS classification, per-corpus freshness dates, and sponsor relationships. Takes only company_id (from find_company). Returns name, ticker, sector/industry classification, country and region, plus a freshness map: latest_intel_date, latest_sec_filings_date (newest 8-K/10-K/10-Q/proxy filing), latest_transcripts_date, and latest_court_filings_date (newest tracked court-case filing). Each freshness field is null when none exists; all dates are ISO-8601 UTC. The response also states each corpus's earliest available date (intel reaches back to 2013-01-01) — freshness is the newest document, coverage is the oldest, and document search defaults to a narrower recent window than full coverage. Also returns sponsors: a list of {sponsor_company_id, sponsor_name, start_date, end_date, status} — the company's Octus-tracked sponsor relationships, historical and current. Active relationships have end_date: null; entries are sorted current-first. sponsors is [] when there are none, and null if the sponsor source was unreachable — do not conflate the two. Coverage is Credit Cloud sponsors coverage, primarily LBO-sourced, not exhaustive. sponsor_company_id is a company id — chain it into get_company_details or list_companies' sponsor param to go deeper. Use this tool to orient before fetching detail with the other tools.
get_deal_term_analytics_data
ChatGPTExtract Deal Term Analytics (DTA) field values for one company's accessible deals. Pass the company_id from find_company and fields — natural-language deal-term descriptions (e.g. ["opening first-lien leverage", "covenant lite", "MFN threshold"]); each is resolved to the DTA catalog internally. Returns metric-major JSON: one entry per resolved field with values keyed by deal (<tx_id>, instrument-level <tx_id>__<instrument_id>); a resolved field with no data still appears (empty values), and asks that match no catalog field are named in note. No server-side math — do any averaging/ranking yourself. Use get_deal_term_analytics_deals first to see which deals are in scope. The response may include a one-row results index carrying the Deal Term Analytics url (individual metrics have no links) — when your answer uses this data, cite inline as a markdown link titled with that row's title, pointing at its url; never construct URLs.
get_deal_term_analytics_deals
ChatGPTList the Deal Term Analytics (DTA) deals your account can access — the orientation card to call before get_deal_term_analytics_data. Each deal carries company_id, company_name, instrument labels (instrument_type, std_instrument, doc_instrument), sponsor, nature_of_deal, doc_date, and an access label (My DTA Deal vs Public DTA Deal) so you can frame 'our deal vs. the market'. Pass company_id (from find_company) to narrow to one company; omit it to list ALL accessible deals. The response may include a one-row results index carrying the Deal Term Analytics url (the deals themselves have no individual links) — when your answer uses this data, cite inline as a markdown link titled with that row's title, pointing at its url; never construct URLs.
get_fundamentals_cap_structure
ChatGPTGet a company's capital structure — the instrument-level debt stack (individual facilities and notes grouped into ranking buckets such as super-senior / senior secured / unsecured / subordinated, each with maturities, coupon/reference rates, and amounts), plus the surrounding analytics: gross and net leverage multiples, liquidity (cash + revolver availability), reconciliations, off-balance-sheet and other liabilities, and hedging. Values are lossless strings, each a {pre, post} pair — display post, falling back to pre only when post is null; net_multiple and gross_multiple are the same shape and follow the same rule. Pass the company_id from find_company. fields is optional: omit it to get the whole table (the natural "show me X's capital structure" ask); pass fields to scope to specific lines, in natural phrasing — each is matched to this company's own line items and the output echoes which line matched. A field may also be a section or category ("liquidity", "senior secured debt", "leverage") — it fans out to all member lines. Unmatched fields return available_fields, this model's full line-item dictionary grouped by section; re-call with exact labels from it instead of rephrasing. Use this for the debt/capital structure and the leverage/liquidity metrics around it; use get_fundamentals_standard_data for Octus-normalized cross-company metrics, get_fundamentals_reported_data for as-filed line items straight from the company's filings, and search_fundamentals with document_types [5] to find capital-structure documents (the source write-ups) rather than the structured table. include_instruments (default true) merges per-instrument metadata (rates, currency, amortization, ranking) onto each instrument line; set it false to drop that detail. By default returns the most recent periods that have data. To target specific periods, pass filter with filter.period.$in — a list of fiscal periods (YYYYMM / YYYYQN / YYYY), e.g. {"period": {"$in": ["2024Q3", "2024"]}}. A company may return several model blocks; each carries status ("Ongoing" / "Suspended") — answer from the Ongoing block with the latest latest_period unless the user asks about the other model. To render as a table: the line name first, then one column per period, newest first in periods order, then exactly four trailing columns — Maturity, Rate, Net Multiple, Gross Multiple — ALWAYS LAST and in that order, in every section (the line's own keys arrive in that order too). Add no other column: each section is one table whose header already names the ranking bucket, so the bucket is a row group, never a column. Each period in values holds one or both VIEWS — actual and pro_forma — and they are different numbers, so render a period that has both as TWO adjacent columns headed "<period> Actual" and "<period> Pro Forma", spelled out in full. A line may carry pro_forma alone (a facility created by the transaction) or actual alone; an absent view is an empty cell meaning not restated, never zero. State the block's currency, but do not state a magnitude or scale: the payload carries none. Results come back as ONE ROW PER MODEL, title naming the company and the model ("<company> (<model type>)"). When your answer uses a figure, cite inline as a markdown link pointing at that row's url, titled with the row's title PLUS the line and period you are citing — e.g. "Alloheim (Public) - RCF - 2026Q1". One row spans the whole table and every period you asked for, so only you know which figure a claim used; never construct URLs.
get_fundamentals_reported_data
ChatGPTGet reported financial values for a company — the figures exactly as the company stated them in its own filings (verbatim, company-specific line items like "Revenue [#2]" or "Senior secured term loan facility, due April 2028"), as opposed to get_fundamentals_standard_data's Octus-normalized lines. Pass the company_id from find_company and fields — the metrics you want, in natural phrasing; each is matched to the company's own reported line items and the output echoes which line matched. A field may also be a category ("KPIs", "income statement items") — it expands to all member lines. Unmatched fields return available_fields, the company's full line-item catalog; re-call with exact labels from it instead of rephrasing. Use this when the user asks for as-reported / as-filed numbers, company-specific line items, or figures the standard set doesn't carry; for cross-company comparisons or normalized metrics use get_fundamentals_standard_data instead (reported labels are not comparable across companies, so this tool is single-company only). By default returns the most recent periods that have data. To target specific periods, pass filter with filter.period.$in — a list of fiscal periods (YYYYMM / YYYYQN / YYYY), e.g. {"period": {"$in": ["2024Q3", "2024"]}}. Results come back as ONE ROW PER MODEL: each row carries title ("<company> (<model type>)"), the model context, and metrics — one entry per metric with its values keyed by fiscal period. When your answer uses a figure, cite inline as a markdown link pointing at that row's url, titled with the row's title PLUS the metric and period you are citing — e.g. "Alloheim (Public) - Revenue - 2025Q3". One row spans every metric and period you asked for, so only you know which figure a claim used; never construct URLs.
get_fundamentals_standard_data
ChatGPTGet standard financial values for a company. Pass the company_id from find_company and fields — the financial line items you want, which you derive from the question: • Specific ask ("what's their revenue and EBITDA?") → pass just those: ["revenue", "EBITDA"]. • Broad 'financials' / 'financial report' request → expand to the standard set (these are Octus's exact canonical names — prefer them; they resolve instantly): ["Revenue", "Revenue by segment", "Gross Profit", "adj. EBITDA", "adj. EBITDA margin", "Net Income", "Gross Debt", "Net Leverage", "Interest Coverage / FCCR", "Cash & Cash Equivalents", "Free Cash Flow"] (growth/trend asks: add YoY lines, e.g. "Revenue Growth YoY"). Any natural phrasing still works — fields resolve against Octus's standard vocabulary, and generic asks ("EBITDA") may fan out to the matching family. Derived metrics include a formula written in standard line names. Unmatched fields return available_fields, the standard line-item catalog grouped by category and correct for that company's model — some lines exist in the vocabulary but are not available on a given model; the note says so explicitly. Re-call with terms closer to those names. By default returns the most recent periods that have data. To target specific periods, pass filter with filter.period.$in — a list of fiscal periods (YYYYMM / YYYYQN / YYYY), e.g. {"period": {"$in": ["2024Q3", "2024"]}}. This tool is standard FINANCIAL data only — covenant/legal/news reports come from their own tools. Results come back as ONE ROW PER MODEL: each row carries title ("<company> (<model type>)"), the model context, and metrics — one entry per metric with its values keyed by fiscal period. When your answer uses a figure, cite inline as a markdown link pointing at that row's url, titled with the row's title PLUS the metric and period you are citing — e.g. "Alloheim (Public) - Revenue - 2025Q3". One row spans every metric and period you asked for, so only you know which figure a claim used; never construct URLs.
list_available_tools
ChatGPTShow all CreditAI search tools and which ones your account can access. Returns: - policy: batch-call policy header — read this first; it states why you must pair general-tier and specialized tools in the same batch. - tools: every tool with has_access indicating whether your account is licensed for it. Each entry returns the same description the tool advertises in its MCP registration plus a short title and tier (general or specialized). - notice (only when present): the account is not enabled for the Octus MCP — every tool reports has_access: false and search calls will be denied. Relay this message to the user instead of retrying. When the user's question is vague or could plausibly fit multiple sources, call this tool and present the accessible tools to the user as a multiple-choice picker (your client's own user-input UI) so they can pick one or more sources before you run a search. Use a per-source tool when a single source is picked, when the user picks several or asks for breadth, run the picked tools in sequential waves of two (the 2-in-flight concurrency cap). Do not call this on every turn — only when disambiguation is genuinely needed.
list_companies
ChatGPTList the companies in a classification / geography bucket — each row is a company_id + company_name (+ any requested fields). Filter by industry_group, industry, and/or region — these are numeric codes (the per-field code=name legend is in each param's schema description; pick the code for the value you want), AND-ed together. Optionally restrict to companies with coverage in one or more products (also numeric codes — see the legend in the schema); multiple products are OR-ed. At least one filter (or product) is required. Results are capped at 100 — if more match, the response says so and you should narrow the search (add/refine filters), not expect more. Also supports sponsor — pass a sponsor firm's company id to get only companies that sponsor backs (current AND former portfolio); AND-ed with the other filters but sufficient on its own. Example: companies backed by KKR — resolve KKR with find_company, then pass its id as sponsor. Pass a returned company_id to get_company_details or the other tools to go deeper.
search_court_filings
ChatGPTSearch Court Filings on CreditAI. Tier: GENERAL — resilience layer. Include at least one general-tier tool in every parallel batch; it's the fallback when specialized sources error or return empty (see server instructions). AI-summarized court filings from US district courts. Good for case-level questions and litigation context. If the user's question is vague or could plausibly fit multiple sources, call list_available_tools first and present the accessible tools to the user as a multiple-choice picker so they can pick one or more — do not guess. Always surface sources for every answer, regardless of format (list, narrative, comparison, memo) — never drop citations. Rendering: answer is a markdown string with inline [N] markers (1-indexed) and a trailing **Sources** block listing [N] [title](url) sourceSubType for each cited source. When presenting the answer to the user, build an N→(title, url) mapping from the **Sources** block, then replace each [N] marker in the prose with [title](url) (a markdown hyperlink using the title as link text). Omit the **Sources** block from the final display — the inline hyperlinks carry the same information. If a footnote line has no markdown link (URL resolution failed), render just the title text in place of [N].
search_covenants_private
ChatGPTSearch Covenants Private Docs and Reports on CreditAI. Tier: SPECIALIZED — narrower scope, may error or return empty. Must be batched with ≥1 GENERAL-tier tool as the fallback (see server instructions). Covenant analysis over private-credit debt documents accessible to the user (sourced from their FinDox data rooms and Deal Term Analytics coverage). Good for the same covenant questions as the public tool — clause comparison, precedent retrieval, provision history, liability-management risk — applied to private-credit deals. If the user's question is vague or could plausibly fit multiple sources, call list_available_tools first and present the accessible tools to the user as a multiple-choice picker so they can pick one or more — do not guess. Always surface sources for every answer, regardless of format (list, narrative, comparison, memo) — never drop citations. Rendering: answer is a markdown string with inline [N] markers (1-indexed) and a trailing **Sources** block listing [N] [title](url) sourceSubType for each cited source. When presenting the answer to the user, build an N→(title, url) mapping from the **Sources** block, then replace each [N] marker in the prose with [title](url) (a markdown hyperlink using the title as link text). Omit the **Sources** block from the final display — the inline hyperlinks carry the same information. If a footnote line has no markdown link (URL resolution failed), render just the title text in place of [N].
search_covenants_public
ChatGPTSearch Covenants Public Docs and Reports on CreditAI. Tier: SPECIALIZED — narrower scope, may error or return empty. Must be batched with ≥1 GENERAL-tier tool as the fallback (see server instructions). Covenant analysis over public-side debt documents — credit agreements, indentures, offering memorandums, and tear sheets for syndicated loans and high-yield bonds. Good for covenant questions: comparing covenant packages, retrieving precedent clauses, understanding provision meaning and history, and spotting liability-management exposure (uptiering, drop-downs, etc.) on public credits. If the user's question is vague or could plausibly fit multiple sources, call list_available_tools first and present the accessible tools to the user as a multiple-choice picker so they can pick one or more — do not guess. Always surface sources for every answer, regardless of format (list, narrative, comparison, memo) — never drop citations. Rendering: answer is a markdown string with inline [N] markers (1-indexed) and a trailing **Sources** block listing [N] [title](url) sourceSubType for each cited source. When presenting the answer to the user, build an N→(title, url) mapping from the **Sources** block, then replace each [N] marker in the prose with [title](url) (a markdown hyperlink using the title as link text). Omit the **Sources** block from the final display — the inline hyperlinks carry the same information. If a footnote line has no markdown link (URL resolution failed), render just the title text in place of [N].
search_data_rooms
ChatGPTSearch Data Room Docs on CreditAI. Tier: SPECIALIZED — narrower scope, may error or return empty. Must be batched with ≥1 GENERAL-tier tool as the fallback (see server instructions). FinDox data rooms. Holds a wide range of materials per deal (analysis, earnings, transcripts, memos, legal docs, etc.). Good for any question grounded in the user's accessible deal data rooms. If the user's question is vague or could plausibly fit multiple sources, call list_available_tools first and present the accessible tools to the user as a multiple-choice picker so they can pick one or more — do not guess. Always surface sources for every answer, regardless of format (list, narrative, comparison, memo) — never drop citations. Rendering: answer is a markdown string with inline [N] markers (1-indexed) and a trailing **Sources** block listing [N] [title](url) sourceSubType for each cited source. When presenting the answer to the user, build an N→(title, url) mapping from the **Sources** block, then replace each [N] marker in the prose with [title](url) (a markdown hyperlink using the title as link text). Omit the **Sources** block from the final display — the inline hyperlinks carry the same information. If a footnote line has no markdown link (URL resolution failed), render just the title text in place of [N].
search_deal_term_analytics
ChatGPTSearch Deal Term Analytics on CreditAI. Tier: SPECIALIZED — narrower scope, may error or return empty. Must be batched with ≥1 GENERAL-tier tool as the fallback (see server instructions). Deal Term Analytics — granular deal-term data extracted from credit documents and benchmarked against Octus's global library, with 170+ data fields per instrument across syndicated loans and private credit. Good for market-wide deal-term comparisons, identifying market-standard vs. outlier terms, screening by industry / sponsor / facility size, and portfolio-level term tracking. If the user's question is vague or could plausibly fit multiple sources, call list_available_tools first and present the accessible tools to the user as a multiple-choice picker so they can pick one or more — do not guess. Always surface sources for every answer, regardless of format (list, narrative, comparison, memo) — never drop citations. Rendering: answer is a markdown string with inline [N] markers (1-indexed) and a trailing **Sources** block listing [N] [title](url) sourceSubType for each cited source. When presenting the answer to the user, build an N→(title, url) mapping from the **Sources** block, then replace each [N] marker in the prose with [title](url) (a markdown hyperlink using the title as link text). Omit the **Sources** block from the final display — the inline hyperlinks carry the same information. If a footnote line has no markdown link (URL resolution failed), render just the title text in place of [N].
search_fundamentals
ChatGPTSearch Fundamentals Data on CreditAI. Tier: SPECIALIZED — narrower scope, may error or return empty. Must be batched with ≥1 GENERAL-tier tool as the fallback (see server instructions). Fundamentals data on companies — both structured and unstructured. Structured side: revenue, EBITDA, leverage, interest coverage, margins, multiples, cash flow, and other reported financials as quarterly / annual time series. Unstructured side: company-disclosed documents including earnings presentations and reports, offerings / lender memoranda, capital-structure docs, press releases, sustainability reports, transcripts and transcript summaries, and compliance certificates. Good for quantitative questions, benchmarking, trend lookups, or anything grounded in these company documents. If the user's question is vague or could plausibly fit multiple sources, call list_available_tools first and present the accessible tools to the user as a multiple-choice picker so they can pick one or more — do not guess. Always surface sources for every answer, regardless of format (list, narrative, comparison, memo) — never drop citations. Rendering: answer is a markdown string with inline [N] markers (1-indexed) and a trailing **Sources** block listing [N] [title](url) sourceSubType for each cited source. When presenting the answer to the user, build an N→(title, url) mapping from the **Sources** block, then replace each [N] marker in the prose with [title](url) (a markdown hyperlink using the title as link text). Omit the **Sources** block from the final display — the inline hyperlinks carry the same information. If a footnote line has no markdown link (URL resolution failed), render just the title text in place of [N].
search_intel
ChatGPTSearch Octus Intel on CreditAI. Tier: GENERAL — resilience layer. Include at least one general-tier tool in every parallel batch; it's the fallback when specialized sources error or return empty (see server instructions). Analyst-written articles, notes, and reports across Octus's editorial coverage: Covenants Analysis — in-depth analysis on US/EU bonds and loans; Credit Intelligence — insights across the entire sub-investment-grade universe; First Day Intelligence — analysis when new Chapter 11 cases are filed; Middle Market Intelligence — actionable intelligence on middle-market trends and transactions; Municipals Intelligence — insights on high-yield municipal issuers and credit trends; Private Company Analysis — in-depth research into private companies across the credit lifecycle; Private Credit & Deal Origination — private-credit and origination intelligence; Structured Finance — CLO market news, manager performance, and portfolio analytics. Use this for the analyst narrative or commentary on any of these areas. Use a source-specific tool instead when the user wants the underlying document. If the user's question is vague or could plausibly fit multiple sources, call list_available_tools first and present the accessible tools to the user as a multiple-choice picker so they can pick one or more — do not guess. Always surface sources for every answer, regardless of format (list, narrative, comparison, memo) — never drop citations. Rendering: answer is a markdown string with inline [N] markers (1-indexed) and a trailing **Sources** block listing [N] [title](url) sourceSubType for each cited source. When presenting the answer to the user, build an N→(title, url) mapping from the **Sources** block, then replace each [N] marker in the prose with [title](url) (a markdown hyperlink using the title as link text). Omit the **Sources** block from the final display — the inline hyperlinks carry the same information. If a footnote line has no markdown link (URL resolution failed), render just the title text in place of [N].
search_my_uploads
ChatGPTSearch My Uploads on CreditAI. Tier: SPECIALIZED — narrower scope, may error or return empty. Must be batched with ≥1 GENERAL-tier tool as the fallback (see server instructions). Search across documents the user has uploaded to CreditAI. Good for any question grounded in their own uploaded files. If the user's question is vague or could plausibly fit multiple sources, call list_available_tools first and present the accessible tools to the user as a multiple-choice picker so they can pick one or more — do not guess. Always surface sources for every answer, regardless of format (list, narrative, comparison, memo) — never drop citations. Rendering: answer is a markdown string with inline [N] markers (1-indexed) and a trailing **Sources** block listing [N] [title](url) sourceSubType for each cited source. When presenting the answer to the user, build an N→(title, url) mapping from the **Sources** block, then replace each [N] marker in the prose with [title](url) (a markdown hyperlink using the title as link text). Omit the **Sources** block from the final display — the inline hyperlinks carry the same information. If a footnote line has no markdown link (URL resolution failed), render just the title text in place of [N].
search_private_company_analysis
ChatGPTSearch Private Company Analysis on CreditAI. Tier: SPECIALIZED — narrower scope, may error or return empty. Must be batched with ≥1 GENERAL-tier tool as the fallback (see server instructions). Deep analysis on private companies, with coverage spanning broadly syndicated loans and high-yield bonds, developed by Octus's experts. Good for in-depth questions on a specific private name. If the user's question is vague or could plausibly fit multiple sources, call list_available_tools first and present the accessible tools to the user as a multiple-choice picker so they can pick one or more — do not guess. Always surface sources for every answer, regardless of format (list, narrative, comparison, memo) — never drop citations. Rendering: answer is a markdown string with inline [N] markers (1-indexed) and a trailing **Sources** block listing [N] [title](url) sourceSubType for each cited source. When presenting the answer to the user, build an N→(title, url) mapping from the **Sources** block, then replace each [N] marker in the prose with [title](url) (a markdown hyperlink using the title as link text). Omit the **Sources** block from the final display — the inline hyperlinks carry the same information. If a footnote line has no markdown link (URL resolution failed), render just the title text in place of [N].
search_sec_filings
ChatGPTSearch SEC Filings on CreditAI. Tier: GENERAL — resilience layer. Include at least one general-tier tool in every parallel batch; it's the fallback when specialized sources error or return empty (see server instructions). SEC filings — 8-K, 10-K, 10-Q, and DEF 14A. Good for questions grounded in what companies have disclosed. If the user's question is vague or could plausibly fit multiple sources, call list_available_tools first and present the accessible tools to the user as a multiple-choice picker so they can pick one or more — do not guess. Always surface sources for every answer, regardless of format (list, narrative, comparison, memo) — never drop citations. Rendering: answer is a markdown string with inline [N] markers (1-indexed) and a trailing **Sources** block listing [N] [title](url) sourceSubType for each cited source. When presenting the answer to the user, build an N→(title, url) mapping from the **Sources** block, then replace each [N] marker in the prose with [title](url) (a markdown hyperlink using the title as link text). Omit the **Sources** block from the final display — the inline hyperlinks carry the same information. If a footnote line has no markdown link (URL resolution failed), render just the title text in place of [N].
search_structured_datasource
ChatGPTAnswer a natural-language question against a structured datasource over Octus data. Pick the datasource code and ask in plain English; the engine works out the data model, runs the query, and returns structured datasets (columns + rows) with the applied period/filters. Read-only; no company_id needed. datasource codes: • 1 (advisors) — advisor engagements over Octus Screener data: who advised which bankrupt / distressed companies (law firms, financial advisors, claims & noticing agents), appointment side (debtor vs. creditor / committee), roles, engagement fees, hourly-rate bands, and league tables by sector / jurisdiction / year, with engine-side grouping, binding-gated aggregation, and negative-existence screens. E.g. 'top financial advisors to debtors in 2023 retail bankruptcies by number of engagements', 'average partner hourly rate for lead debtor counsel'. • 2 (primary_tracker) — LevFin new-issue primary market: leveraged-loan and high-yield-bond tranches with launch / pricing dates, initial and final sizes, coupon / margin / yield / OID economics, price talk, ratings, bookrunners, sponsors, and use of proceeds, with engine-side counts, league tables, grouped averages / sums, and per-year / quarter time series. E.g. 'how many high-yield bonds priced in 2025', 'top bookrunners for 2025 leveraged loans', 'average TLB margin by market of risk'. (More datasource codes will be added as they productize.) Returns datasets — one entry per query the engine ran, each a {label, columns, rows, ...} table (rows are {column: value}) — plus an optional one-line summary and an optional note. When the question can't be answered (off-topic, a metric not in the model, or an aggregation the data can't support), the tool does not error — it returns an empty datasets list with a note explaining why in business terms, usually with a workable rephrasing; a partial answer likewise comes back with data plus a note. Always read the note and, when present, re-ask using its guidance. Rows are read-only structured values, not document excerpts — present the tables and any rankings directly. The response may include a one-row results index carrying the url of the Octus page this datasource's data lives on (a screen is a query, not a document, so individual rows have no links of their own) — when your answer uses this data, cite inline as a markdown link titled with that row's title, pointing at its url; never construct URLs.
search_transcripts
ChatGPTSearch Transcripts on CreditAI. Tier: GENERAL — resilience layer. Include at least one general-tier tool in every parallel batch; it's the fallback when specialized sources error or return empty (see server instructions). Company transcripts including earnings calls and other transcribed management events. Good for what management actually said, direct quotes, and call-level detail. If the user's question is vague or could plausibly fit multiple sources, call list_available_tools first and present the accessible tools to the user as a multiple-choice picker so they can pick one or more — do not guess. Always surface sources for every answer, regardless of format (list, narrative, comparison, memo) — never drop citations. Rendering: answer is a markdown string with inline [N] markers (1-indexed) and a trailing **Sources** block listing [N] [title](url) sourceSubType for each cited source. When presenting the answer to the user, build an N→(title, url) mapping from the **Sources** block, then replace each [N] marker in the prose with [title](url) (a markdown hyperlink using the title as link text). Omit the **Sources** block from the final display — the inline hyperlinks carry the same information. If a footnote line has no markdown link (URL resolution failed), render just the title text in place of [N].
describe_tool
Claudefind_company
Claudeget_company_details
Claudeget_deal_term_analytics_data
Claudeget_deal_term_analytics_deals
Claudeget_fundamentals_cap_structure
Claudeget_fundamentals_reported_data
Claudeget_fundamentals_standard_data
Claudelist_available_tools
Claudelist_companies
Claudesearch_court_filings
Claudesearch_covenants_private
Claudesearch_covenants_public
Claudesearch_data_rooms
Claudesearch_fundamentals
Claudesearch_intel
Claudesearch_my_uploads
Claudesearch_private_company_analysis
Claudesearch_sec_filings
Claudesearch_structured_datasource
Claudesearch_transcripts
Claude
