Skills & Tools

​

GitHub

skills: [Skills.github({ repos: [Repos.MyOrg.MyRepo] })]

​

Slack

skills: [Skills.slack({ channel: SlackChannel.Engineering })]

​

Linear

skills: [Skills.linear()]

​

Attio

skills: [Skills.attio({ object: AttioObject.People })]

​

Notion

import { NotionDatabase, NotionPage, Skills } from "./terse.generated"

skills: [Skills.notion({ databases: [NotionDatabase.Roadmap], pages: [NotionPage.OnCallRunbook] })]

​

Gmail

skills: [Skills.gmail()]

​

Gmail Draft

skills: [Skills.gmailDraft()]

​

Meta Ads

import { MetaAdsAdAccount, Skills } from "./terse.generated"

skills: [Skills.metaAds({ adAccount: MetaAdsAdAccount.Primary, pageId: "1234567890" })]

• Each list tool returns its own collection, a count, and truncated.
• Insights need level: "ad" to attribute results to an individual creative, optionally split by breakdowns such as age or publisher_platform. meta_ads_read_insights pages up to 2000 rows and sets truncated: true beyond that; each breakdown multiplies the row count.
• Meta cannot edit a creative or swap the one on an ad. Improve one by creating a new ad in the same ad set and pausing the old one; each ad keeps its own insights.
• New ads enter PENDING_REVIEW and can come back DISAPPROVED. Re-read effective_status with meta_ads_list_ads before treating an ad as live.
• meta_ads_set_status only pauses and resumes. There is no budget tool, so an automation cannot increase spend.

​

Creative formats

cards: [
    { media: "image", imageUrl, headline, description?, linkUrl? },
    { media: "video", videoUrl, headline, description?, linkUrl? }
]

• Only the fields you supply are sent, so an omitted callToAction leaves Meta’s default button. status is the exception: it defaults to PAUSED, not Meta’s ACTIVE.
• Carousels containing a video card are Facebook-only: Meta does not support them on Instagram placements.
• Dynamic creative is capped at 30 assets across media and text (ad_format and link_url each count as one), requires callToActions, and only works on an ad set with dynamic creative already enabled in Ads Manager.
• Images are fetched by Meta from the URL you supply; videos and dynamic-creative images are uploaded first, so short-lived signed URLs work everywhere, including the ones Higgsfield returns. The result reports videoIds and imageHashes.

​

Connecting

​

Higgsfield

import { Skills } from "./terse.generated"

skills: [Skills.higgsfield()]

• The tools block until generation finishes. Each asset is copied into Terse storage and returned as a signed URL valid for 24 hours, ready to pass to meta_ads_create_ad as an imageUrl or videoUrl.
• Every call spends Higgsfield credits, so both generators are writes and can be gated behind approval. Use batchSize: 4 for variants to compare, and match size to the placement (2048x1152 for feed, 1152x2048 for stories).
• model trades cost against quality: dop-lite, dop-turbo (default), dop-standard.
• A motion is a named camera move applied on top of your prompt. IDs are opaque, so discover them with higgsfield_list_motions, pass one as motionId, and scale it with motionStrength (0-1, defaults to 1).

const { images } = await toolbox.higgsfield.generateImage({
    prompt: "A coffee tin on concrete, soft daylight, editorial styling",
    size: "1152x2048"
})
const { videos } = await toolbox.higgsfield.generateVideo({
    imageUrl: images[0].url,
    prompt: "Slow push in, steam rising, shallow depth of field",
    model: "dop-turbo"
})

​

Connecting

terse integrate connect higgsfield --fields-stdin <<<'{"credentials":"KEY_ID:KEY_SECRET"}'

​

Snowflake

skills: [Skills.snowflake()]

​

Datadog

import { DatadogIndex, Skills } from "./terse.generated"

skills: [Skills.datadog({ indexes: [DatadogIndex.Main] })]

​

PostHog

import { PosthogProject, Skills } from "./terse.generated"

skills: [Skills.posthog({ project: PosthogProject.MyTeam })]

​

LaunchDarkly

import { LaunchDarklyProject, Skills } from "./terse.generated"

skills: [Skills.launchDarkly({ project: LaunchDarklyProject.MyFlags, environmentKeys: ["production"] })]

​

WorkOS

skills: [Skills.workOS()]

​

Apollo

skills: [Skills.apollo()]

• Person and organization enrichment consume Apollo export credits per matched record. Prefer apolloBulkEnrichPeople for lists.
• apolloSearchPeople consumes no credits but requires the connected key to be an Apollo master API key, and its results never include emails — pass result ids to apolloBulkEnrichPeople to unlock contact data.
• apollo_list_job_postings works with any Apollo key but consumes credits per page returned — fetch one large page (up to 500 postings) instead of paging in small steps. Postings carry title/URL/location/date metadata only, no descriptions.
• Personal emails are only returned when revealPersonalEmails is set, and Apollo suppresses them for people in GDPR regions.

​

Google Search Console

skills: [Skills.googleSearchConsole({ sites: [GoogleSearchConsoleSite.DomainAcmeCom] })]

• A property is either a URL-prefix property (https://example.com/, trailing slash included) or a Domain property (sc-domain:example.com).
• The properties you pass to the skill are the only ones the agent may act on, except google_search_console_list_sites, which lists everything the account can see. A Domain property also covers its subdomains.
• Search Analytics dates are inclusive YYYY-MM-DD in PST and the data lags by roughly 2-3 days, so the most recent days come back empty rather than as zero traffic.
• Query-grouped results omit anonymized queries, so their clicks will not sum to the property total.
• Submitting a sitemap only queues it. Check the outcome later with google_search_console_get_sitemap.
• Property access follows the connected Google account, not your Terse organization, so connect the account that actually appears in Search Console.

​

Web (built-in)

skills: [Skills.web()]

skills: [Skills.web({ allowedDomains: ["example.com", "docs.acme.com"] })]

​

Image Edit (built-in)

skills: [Skills.imageEdit()]

​

Memory (built-in)

skills: [Skills.memory()]

​

What persists between runs

​

How memory is organized

• Per project — each project has its own isolated memory. Jobs in different projects never share memory.
• Per job — within a project, every job gets its own private space; one job cannot see another job’s memory.
• Scoped to /memories — all paths stay inside the /memories directory.

​

Inspecting memory