"How to increase a model's context window for handling large files" matching MCP tools:

• get_dataset_filesA

  gdex-mcpA

  Return a paginated file listing for a dataset or a specific group. Depending on how deep gindex is in the group hierarchy, this returns either actual file rows or another layer of subgroup summaries — there's no way to tell in advance which you'll get. If you get subgroups, read the gindex (or url) off a row and call again with that gindex to go one level deeper; gindex values are dataset-specific and can't be guessed. A shallow gindex on a large dataset can return a very large response (thousands of files) — the file-row portion of the response is capped at 500 rows (look for "_truncated": true). Two ways to avoid hitting that cap instead of drilling down group by group: pass filter_wfile with a filename pattern (e.g. a date like "20220808") to filter down to matching files, or page through a known group's results with `page`. filter_wfile only filters actual file rows, so it has no effect at a gindex that's still returning a subgroup summary rather than files — if a first attempt comes back unfiltered, descend one level (see get_file_groups) and retry there. Args: dsid: Dataset ID (dNNNNNN), e.g. d083002 gindex: Optional group index to filter files page: Page number to fetch (for a group with more files than fit on one page) filter_wfile: Filter files by name pattern, e.g. "20220808" to match a date fl: File list source (defaults to "web" server-side)

  ConnectorNo auth

• update_siteA

  vibedeployA

  Patch or replace files on an existing site. Defaults to patch mode: only the listed files change; everything else stays. Pass mode:'replace' to wipe-and-replace the whole site (the legacy behaviour, surfaced explicitly so it can't happen by accident). Use `delete: [paths]` in patch mode to remove specific files without wiping the rest. Use `dryRun: true` to preview the diff before committing. LARGE FILES: a 100-250 KB text file fits in one call with encoding:'gzip+base64' (gzip locally, base64 the result) — prefer that over begin_deploy + add_file_chunk streaming. Errors if the site does not exist.

  Connector

  Destructive

  No auth

• nukez_storeA

  nukez-mcpA

  Store files in your Nukez locker. ALWAYS BATCH: pass ALL the files you want to upload in a SINGLE call, as a list under `files`. Do NOT loop over your file list and call nukez_store once per file — that triggers one on-chain attestation per file (slow + costs SOL fees per push). One nukez_store call with N files triggers exactly ONE attestation for the whole batch. Each list item: {name, <data_source>, content_type?, expected_size_bytes?, expected_sha256?, large?}. UPLOAD PATH PRIORITY: 1. source_url — if the file is available at a public HTTPS URL, pass it and the server fetches directly (fastest, zero token cost). 2. sandbox_path — if you have compute/bash access and the file is on disk, pass the absolute path. 3. local_path — if the file exists on local disk (desktop/CLI environments). 4. data_b64 — LAST RESORT for small content only (<4KB). Sends bytes through your context window. Accepts both base64-encoded binary and plain UTF-8 text. 5. nukez_upload_chunk — if sandbox_path curl failed (HTTP 000 / network blocked), upload in 4KB chunks with sha256 verification. Run the prep script from the response, then call nukez_upload_chunk for each chunk. NEVER base64-encode files >4KB in one call. Upload path is auto-selected based on size and runtime environment. HARD SIZE LIMITS per path: the sandbox curl proxies through the gateway and is capped at 32 MiB (the platform rejects larger request bodies); the chunked ingest pipeline caps at 20 MiB per file. LARGE FILES (>= 32 MiB, any size up to terabytes): pass expected_size_bytes (and ideally the file's expected_sha256) with sandbox_path, or set large=true — the response then returns a resumable direct-to-provider upload session (action_required='execute_upload') with a session-open command and a stdlib-only transfer script that uploads in 8 MiB chunks and resumes after interruptions. Bytes go straight to the storage provider, so the sandbox needs outbound access to the session URI's host. After the transfer, confirm with nukez_confirm(use_job=true). KEYLESS (hosted) SERVER: this server holds no signing key. A call without `envelope` returns action_required='sign_envelopes' with the exact spec to sign (method, path, ops, body); sign it with your wallet and re-call with envelope=<signed result>. One file per call in envelope mode.

  ConnectorNo auth

• get_trip_windowA

  SnowSure — Snow & SkiA

  Historical odds for a SPECIFIC trip window at one resort — 'Whistler on March 27', 'Vail Jan 10–17'. From ~30 years of ERA5 monthly history it blends the window's months (weighted by days) into: the snow-day probability (the headline — for a short window how OFTEN it snows beats how MUCH falls in a season), window-scaled expected snowfall and typical base depth, the same window across the last 5 individual seasons, and the best nearby month (±1) by snow-day odds. Window max 31 days; dates are YYYY-MM-DD. This is HISTORY, not a forecast and not a prediction — for the next 14 days use get_weather_forecast, and for a whole-month question use get_monthly_snow. Windows with under 5 tracked seasons return insufficient rather than thin odds.

  ConnectorNo auth

• search_by_sentimentA

  Islam West Africa Collection (IWAC)A

  Filter articles by gpt-5-6-luna sentiment labels (accent/case-insensitive exact match). One model's reading, not a consensus — 4 other models scored the same articles and often disagree; get_sentiment_distribution with model:"all" shows by how much. `subjectivity` is much the weakest of the three scales, so treat a set selected on it as a lead to read rather than as a finding.

  ConnectorNo auth

• discovery_uploadA

  DiscoA

  Upload a dataset file and return a file reference for use with discovery_analyze. Call this before discovery_analyze. Pass the returned result directly to discovery_analyze as the file_ref argument. Provide exactly one of: file_url, file_path, or file_content. Args: file_url: A publicly accessible http/https URL. The server downloads it directly. Best option for remote datasets. file_path: Absolute path to a local file. Only works when running the MCP server locally (not the hosted version). Streams the file directly — no size limit. file_content: File contents, base64-encoded. For small files when a URL or path isn't available. Limited by the model's context window. file_name: Filename with extension (e.g. "data.csv"), for format detection. Only used with file_content. Default: "data.csv". api_key: Disco API key (disco_...). Optional if DISCOVERY_API_KEY env var is set.

  ConnectorNo auth

• get_connector_clientA

  ValetA

  Get how a page served from a Valet site talks to the connectors attached to it. It is the pair to get_skill with design-system: consult that for the artifact's identity, this one for how it talks to its connectors. Call it before writing any page code that fetches from a connector. Returns the same-origin request contract and a paste-whole session helper that handles both sessionless and stateful connectors, plus the rules a page has to follow: handling a 403, recovering a lapsed session, and never caching a response on the caller's behalf. This needs no Valet account. It answers with static guidance about how pages work — nothing here reads an organization or a site — so an agent can call it before running any OAuth flow.

  ConnectorNo auth

• add_artifactA

  TangoA

  Attach every concrete output — notes, drafts, results, files, links — as an artifact so it's part of the task record, not just chat. Real files (pdf, docx, pptx, xlsx, mp3, wav, m4a, images…) are supported: pass `content_base64` for files up to ~6 MB, `fetch_url` to have Tango download and store a hosted file itself, or call create_artifact_upload first for large files and finalize here with `upload_token`. `content` stays the path for inline text and `external_url` for a link you only want recorded. Reference artifact ids in complete_task's evidence_artifact_ids. If a lease is active, Tango attributes the artifact to the lease holder. Otherwise, pass `acting_worker_id` to identify which of your workers is acting; if you don't own that worker the attribution is dropped rather than misrecorded. Attested workers may pass `worker_signature` over the JCS-canonical artifact payload (type 'tango.artifact'); an invalid signature rejects the call and nothing is stored. Delegated workers are signed for automatically. API reference: https://tango.applayer.io/docs/api/tools/add_artifact

  ConnectorOAuth

• measureA

  3dCMSA

  Measure nodes in a saved scene (analytic, instant, no render). For each node in nodeIds: its world centre, its size along its own axes after every scale above it, its world bounding box, and `clearance`: the nearest visible thing in each world direction (+x, -x, +y, -y, +z, -z) that overlaps it across that axis, with the gap in meters (how far a frame hangs off its wall, how much room a camera has before it meets a side wall). An IMPORTED MODEL also gets `opening`: the hole through it (a picture frame, a doorway, a window, an arch), measured from the model's own triangles: its size in meters along the node's x, y, z, the axis it goes through, and `backed` when a board or glass covers the centre. Size a panel a little larger than the opening so its edge tucks behind the lip; get_scene only reports a model's OUTSIDE. Two or more nodes also give `pairs`: centre distance, the gap between their boxes (0 when they overlap) and that gap per axis. A `point` [x,y,z] gives its clearance in all six directions.

  ConnectorNo auth

• files_get_base64A

  DialogBrainB

  Download one or more files server-side and return their content as base64-encoded strings. Use this to inspect images, PDFs, or any binary file attached to messages when you cannot access presigned S3 URLs directly. Supports up to 5 files per call, max 15 MB each. For large files batch in groups of 1-2 to avoid oversized responses. For images, pass `variants` to get processed copies instead of the original: rotate (clockwise) a photo taken sideways or upside down, enhance a faded print, or crop and enlarge a region with small text. Ask for both rotate 90 and 270 in one call when the text runs vertically and you cannot tell which way is up. An MCP client shows at most 5 images per response; the rest are listed, not shown.

  ConnectorAPI key

• files_get_base64A

  dialogbrainB

  Download one or more files server-side and return their content as base64-encoded strings. Use this to inspect images, PDFs, or any binary file attached to messages when you cannot access presigned S3 URLs directly. Supports up to 5 files per call, max 15 MB each. For large files batch in groups of 1-2 to avoid oversized responses. For images, pass `variants` to get processed copies instead of the original: rotate (clockwise) a photo taken sideways or upside down, enhance a faded print, or crop and enlarge a region with small text. Ask for both rotate 90 and 270 in one call when the text runs vertically and you cannot tell which way is up. An MCP client shows at most 5 images per response; the rest are listed, not shown.

  ConnectorAPI key

• files_get_base64A

  dialogbrainB

  Download one or more files server-side and return their content as base64-encoded strings. Use this to inspect images, PDFs, or any binary file attached to messages when you cannot access presigned S3 URLs directly. Supports up to 5 files per call, max 15 MB each. For large files batch in groups of 1-2 to avoid oversized responses. For images, pass `variants` to get processed copies instead of the original: rotate (clockwise) a photo taken sideways or upside down, enhance a faded print, or crop and enlarge a region with small text. Ask for both rotate 90 and 270 in one call when the text runs vertically and you cannot tell which way is up. An MCP client shows at most 5 images per response; the rest are listed, not shown.

  ConnectorAPI key

• resolve_topicA

  Pine ScriptA

  USE WHEN looking up an exact Pine Script API term or known concept keyword. Returns the best-matching doc paths with matched keywords and a retrieval suggestion (get_doc or list_sections + get_section). AFTER calling this tool, follow the suggestion: call get_doc() for small files or list_sections() + get_section() for large files. For natural language questions use search_docs() instead. Data sourced from bundled TOPIC_MAP and doc file content scan.

  ConnectorNo auth

• get_insider_pulseA

  mcpA

  Get SEC Form 4 insider-trading activity — per-ticker detail or S&P 500 screener. Surfaces already-ingested Form 4 filings (the same Form 4 feed behind the market model's `form4` stage) pre-chewed into analytical context so you don't have to parse raw XBRL yourself: buyer role (officer / director / 10% owner), transaction size vs that SAME insider's own historical buy pattern (`size_vs_own_median`, `is_unusually_large` when >=2x their own median), and cluster-buy detection (`is_cluster` — 3+ distinct insiders buying the same ticker within a 10-day window, historically the stronger signal vs a single insider's trade). Every transaction links to its source SEC filing via `accession_url` — verify anything before acting. Two modes: - ticker set: per-ticker view — up to 40 most-recent Form 4 transactions over the trailing year plus `cluster_buy_active`. - ticker omitted: S&P 500 (top-100) screener — tickers with a cluster buy or an unusually-large buy in the last 30 days. An empty `results` list is a valid, honest answer; most weeks most tickers show nothing notable. Read-only surfacing: this is information, not a recommendation. Args: ticker: Optional. Stock ticker for the per-ticker view (e.g. "AAPL"). Omit for the S&P 500 screener. Case-insensitive. Returns per-ticker: ticker, transactions, cluster_buy_active, generated_at, disclaimer. Screener: tickers_scanned, results, generated_at, disclaimer.

  ConnectorNo auth

• get_boundary_geojson_simplifiedA

  BoundrA

  Fetch simplified GeoJSON for a boundary by its ONS census code. Safe to embed directly in generated HTML map files. At the default tolerance (0.0001°) a constituency polygon shrinks from ~4,000 vertices to ~200–400 with no visible difference at normal map zoom levels. Prefer this over get_boundary_geojson_by_code() when writing Leaflet map pages — the full geometry is large enough to exhaust your context window before you can finish writing the HTML.

  ConnectorNo auth

• get_price_historyA

  The Front Desk ReviewA

  The full dated price series for ONE product or vendor (by slug), plus derived stats: first-tracked date, current/low/high, number of recorded changes, net % since tracked, the largest single move, and the last change. This is the citable answer to 'how much did X cost', 'did X's price drop', and price-trend questions — grounded on a real, append-only series, not a model's guess. Returns found:false (never a fabricated series) for an unknown slug.

  ConnectorNo auth

• get_stateA

  valemA

  Get a model's merged state (base fields plus all computed derived fields). On a large model this is the biggest context cost — narrow it: pass 'paths' to project only specific subtrees (canonical addresses, each spliced back into a pruned document at its address), and/or 'depth' to cap nesting (deeper containers collapse to a '<object: N fields>' / '<array: N items>' marker). Pass an optional ISO-8601 'at' timestamp for a point-in-time read from mutation history.

  ConnectorNo auth

• build_stage_bundleA

  SuperoB

  PUB-1a: get the out-of-band UPLOAD endpoint for a LARGE app bundle that won't fit inline (the model's max OUTPUT tokens cap `files`/`files_b64gz`, so big apps otherwise have to be truncated/minified). Upload a gzip(json {path:content}) blob to the returned URL with your OWN key, then pass the returned file_id as `files_ref` to build_validate / build_publish / build_doctor. Bundle size then no longer depends on any token cap.

  ConnectorNo auth

• share_fileA

  agent-cloud-toolsA

  Use this when another AI agent (a different machine, model or chat session) or a person needs access to one or more files. Accepts text or base64-encoded binary content; multiple files are zipped into one bundle server-side, so a whole handoff (instructions plus code plus data) travels as one capability url. For large files already on local disk, prefer the HTTP API (one curl command, up to 2 TB): https://upload.83blue.com/docs

  ConnectorNo auth

• find_replacementA

  AI WaveA

  What to switch to when a model is gone or you need a fallback. Ranked by closeness in measured performance, not by vendor or price: what you usually need to preserve first is the quality of the output. Candidates whose context window is less than half the original are excluded. Returns the score, price and context deltas so you can judge; we do not pick for you.

  ConnectorNo auth

• get_downloadsA

  Odile LabsA

  Signed download links for every finished file of a job, once unlock_job has succeeded for it. Use it as the last step of the flow, and again whenever links have expired. Needs a signed-in connection and changes nothing. Returns { job_id, files[{name, preset, bytes, width, height, duration, url, expires_at}], how }. Each link works for 15 minutes and needs no login: fetch the files with curl (see how) rather than reading them into the conversation, or, without a shell, give the customer the links to open. Files are kept for 30 days. A job that is not unlocked returns an error with its state and cost; call unlock_job first.

  ConnectorOAuth