agentbull/ankole/context-dev

Context.dev

Context.dev turns a URL, a domain, or file bytes into typed JSON. Every call spends credits from the operator's Context.dev account, so select the narrowest tool that answers the question, and keep a result instead of asking for it a second time.

Select the Owner First

Ankole already owns most of this work. Use the built-in owner when it applies:

Route to Context.dev for the work those owners do not cover:

• A page that refuses an ordinary fetch, or that shows its content only after a wait or a click.
• A site inventory, a whole-site crawl, or an asynchronous batch of up to 25,000 URLs.
• A page or site that must answer in a caller-defined JSON Schema.
• A company profile, design system, font list, page screenshot, or NAICS or SIC code.
• A page, sitemap, or extraction that must be watched on a schedule.

MCP Entry Point

The configured mcporter server is context-dev. Use Tool Selection below to choose one exact tool before you inspect or call the server.

When the arguments are not already fixed by this Skill and its references, inspect only the selected tool:

mcporter list 'context-dev.<tool-name>' --schema --json --timeout 360000

Write the argument object as JSON to a temporary file, then call the tool without shell interpolation. Send the result to an absolute file path and read that file:

mcporter call 'context-dev.<tool-name>' --json - --output json --timeout 360000 \
  < /absolute/path/to/arguments.json > /absolute/path/to/result.json

Read result.json, or reduce it first with jq or python3 when it is large. Remove the argument file and the result file after each call.

mcporter loses most of its own output when stdout is a pipe, and the command tool captures stdout through a pipe. Page Markdown, crawl results, sitemaps, and batch pages are large enough to be cut in the middle of a token with exit code 0 and no error on stderr. Write to a file, and treat a JSON parse failure on mcporter output as this truncation until you prove otherwise.

In an Applet Automation Job, call AnkoleRuntime.mcp.call(tool, arguments, {server: "context-dev", skill: "context-dev"}). The Function Runtime host runs MCPorter and returns the structured result. A personal Job uses its authorizing human; a service Job uses AnkoleRuntime.admin.mcp.call to act for the Page owner.

Ankole generates MCPORTER_CONFIG for each Main turn or Background Agent Job execution from this enabled Skill. The configuration points to the control-plane gateway. The control plane resolves CONTEXT_DEV_API_KEY from the Agent's configured credentials and injects it only into the upstream request. If authentication fails, check the configured credential; the Worker and Applet do not receive the external key.

Tool Selection

Web reading

• web-search: Search the live web. 1 credit for each result, and numResults starts at 10, so one search costs at least 10 credits. markdownOptions.enabled: true also scrapes each hit in the same call.
• web-scrape-markdown: Turn one known URL into clean Markdown. 1 credit, or 2 with actions. This is the default page reader.
• web-scrape-html: Return rendered raw HTML. 1 credit, or 2 with actions. Use it only when DOM structure, attributes, or scripts are the answer.
• web-scrape-images: List the image assets of one page. 1 credit, 2 with actions, 5 when any enrichment flag is set.
• web-scrape-sitemap: Discover the URLs of a domain. 1 credit. It returns URLs only, never page bodies, so use it before you decide what to read.
• web-crawl: Follow internal links from one URL and return Markdown for each page in one synchronous call. 1 credit for each page, maxPages at most 500. Set maxPages to what you will read.
• web-extract: Crawl the relevant pages and return data shaped by a JSON Schema. 10 credits, maxPages at most 50.
• parse-document: Convert base64 file bytes to Markdown. 1 credit, 5 when OCR adds text, and the decoded file must be 25 MiB or smaller. Ankole reads PDF, image, DOCX, PPTX, and XLSX files locally through its own Skills, so reach for this tool only for a format they do not read, such as .doc, .rtf, or .srt.

Read references/web-reading.md before you compose any of these calls.

Brand intelligence

• brand-retrieve-unified: Return a structured company profile. 10 credits. All arguments go inside one body object whose type selects the lookup.
• get-brand: Return the same profile for a bare domain as a visual card. 10 credits. Prefer brand-retrieve-unified when you need the fields.

Design, visual, and classification

• web-styleguide: Return colors, typography, spacing, shadows, and paste-ready component CSS. 10 credits.
• web-fonts: Return the font families of a site with fallbacks and usage share. 5 credits.
• web-screenshot: Return a hosted PNG URL for a rendered page. 5 credits.
• web-naics: Return 2022 NAICS codes for a domain or company name. 10 credits.
• web-sic: Return SIC codes against original_sic or latest_sec. 10 credits.

Read references/brand-and-design.md for the body union, the returned brand object, and the domain/directUrl rule.

Monitors

list-monitors, create-monitor, get-monitor, update-monitor, delete-monitor, run-monitor-now, list-monitor-runs, get-monitor-run, list-monitor-changes, list-changes, get-change, list-account-runs, list-monitor-credit-usage.

A monitor is a recurring check that lives in the Context account and spends credits at each run. Use it for change detection over time, not for one reading.

Batches

submit-batch, list-batches, get-batch, get-batch-results, cancel-batch, delete-batch.

A batch is an asynchronous job for work that is too large for web-scrape-markdown or web-crawl. submit-batch returns an ID, not pages.

Read references/monitors-and-batches.md for both lifecycles.

Account Writes

create-monitor, update-monitor, delete-monitor, run-monitor-now, submit-batch, cancel-batch, and delete-batch change durable state in the operator's Context account, and a monitor keeps spending credits until somebody removes it. Call them for the request the user made, and report the monitor or batch ID that the call returned so the user can find it again.

Treat a submission as non-idempotent. After a timeout or an unclear error, list monitors or batches and read the current state before you retry. submit-batch accepts an Idempotency-Key; reuse the same key when you retry the same submission.

Reading Rules

1. Pass a bare domain, such as stripe.com, where the tool asks for domain. Pass a full http or https URL where the tool asks for url or directUrl.
2. A cached answer returns in under a second. A cold answer takes about 7 seconds at p50 and can reach a minute, so raise timeoutMS toward its 300000 ms maximum for a first look at a slow site, and keep the mcporter --timeout above it.
3. [N entries omitted] or ...[truncated] means the business result is incomplete. Do not analyze it and do not advance a cursor. Ask again with fewer pages, fewer results, or a narrower range.
4. Follow next_cursor while the response says more results exist. A page can close early to stay under its size limit, so trust the cursor instead of counting records.
5. Any field can be absent. Check the fields you need before you build an answer on them.
6. Report the URL or domain you read, the tool you called, and any page that failed or was skipped.

Errors

A tool name that the server rejects means the catalog changed. Inspect the selected tool again and use its current replacement.