Driving ShapelessAI

ShapelessAI runs the user's social presence. You reach it through the shapeless MCP tools this plugin ships (or the shapeless CLI, same surface). Everything is one account, gated by a credential - the OAuth sign-in or an API key - with scopes: read, write, publish. Only publish puts content into the world.

The documentation

Do not guess this API. It is published, and it answers Markdown:

https://shapelessai.com/docs - one page per subject. Append .md to any path (https://shapelessai.com/docs/posts.md) or send Accept: text/markdown to get the source instead of the page.
https://shapelessai.com/llms-full.txt - every page in one fetch. Read this when you need the whole surface; read a single .md page when you need one answer.
https://shapelessai.com/api/openapi.json - OpenAPI 3.1 for every route.

Straight to the page you need: posts · platforms · api · jobs · mcp tools · brand memory · auth.

First contact

Call me to confirm the credential works and see the plan and credit balance. If it fails, the user signs in over OAuth: run /mcp, pick shapeless, choose Authenticate, and allow in the browser tab. The stdio fallback (npx shapelessai mcp) needs a key instead: minted at https://shapelessai.com/studio/api-keys, then shapeless login or export SHAPELESS_API_KEY=slk_....

No me tool at all? Then this skill arrived without its server (a ClawHub or skills.sh install carries only this file). The server is hosted at https://shapelessai.com/mcp (Streamable HTTP, OAuth, nothing to run locally). Add it once, then sign in; the browser tab that opens also creates the account when the user has none, and the Free plan is enough to post:

# OpenClaw
openclaw mcp set shapeless '{"url":"https://shapelessai.com/mcp","transport":"streamable-http","auth":"oauth"}'
openclaw mcp login shapeless

# Claude Code
claude mcp add --transport http --scope user shapeless https://shapelessai.com/mcp   # then /mcp, Authenticate

# Codex
codex mcp add shapeless --url https://shapelessai.com/mcp && codex mcp login shapeless

Every other host (Hermes, Cursor, Gemini CLI, ChatGPT, Claude, Grok, Perplexity) has its steps at https://shapelessai.com/connect. Tell the user which command you ran and that one sign-in is theirs to do; the tools appear once the host reloads its servers (the next turn or session).

The shape of the work

Jobs are durable server-side runs: jobs_create with a goal ("draft three posts about the beta launch") returns an id and keeps going whether or not you stay attached. jobs_list lists them, jobs_get shows the transcript and outputs, jobs_brief is the compact version for continuing, jobs_tail watches a run live, jobs_run presses Run on a proposed plan, jobs_continue resumes a stuck or finished run with new instructions, and jobs_stop stops one. Prefer a job for anything generative - posts, carousels, images, narrated video - the server holds the account's Brand Memory, connected platforms, and media pipeline.
Media rides the message as mediaKeys: anything already in the account (assets_list). Up to 6 per message. The agent reads them for real - an image's pixels, not just its filename. The hosted server has no disk, so a local file goes in first with the CLI (shapeless assets upload ./file) or PUT /api/brain/assets/{filename}, which returns the key.
Posts are the queue: proposed -> scheduled -> published. posts_list / posts_get to review, posts_approve to accept a proposal into the schedule, posts_dismiss to reject, posts_publish to put one out now.

posts_create puts a post you wrote on a real account. Call platforms_list first - it carries each platform's text limit, media rules, whether a title is required, first-comment support, and the JSON Schema for its settings. Then:

{
  "connectionId": "<from connections_list>",
  "text": "The post.",
  "scheduledAt": "2026-09-21T09:00:00+03:00",  // omit for now
  "queue": true,                    // instead of scheduledAt: the next free slot
  "mediaKeys": ["..."],             // media already in the account
  "title": "The video title",       // YouTube requires one
  "settings": { },                  // per platform; platforms_list has the schema
  "firstComment": "Link: https://..."   // LinkedIn, X, Bluesky only
}

scheduledAt and queue are exclusive. A platform rule refuses at creation with a 422 that names the rule, never at publish time. The CLI equivalent is shapeless posts create --to <id> --text "..." [--at <ISO> | --queue] [--media k1,k2] [--title "..."] [--settings '{json}'] [--first-comment "..."].

The Free plan posts ten a day on the rail, counted on the UTC day each post goes out on - so a week planned ahead is ten a day, not ten in total, and approving a proposal counts the same as creating one. A post that failed keeps its slot: retry it with posts_publish on the same id, which takes no new slot, never by creating it again. The eleventh answers 402 {code: "free_daily_cap", limit, day, resetsAt} naming the day that is full: move the post to a day with room, or tell the user their account is on Free. X is the exception, because X bills per post: Free posts to X once a day (a first comment counts) and never with a link, in the post or its first comment. A second X post that day answers 402 free_x_daily_cap, a link answers 402 free_x_link; take the link out or move the post to the next day. The Scheduler plan lifts the daily cap and carries 150 X posts a month (a post with a link counts as 10, 402 x_monthly_cap when they run out); the agent plans (Spark, Growth, Studio) have no cap at all. Every 402 carries resetsAt and a checkout link for the human. Composing, scheduling and publishing never spend credits; Free also carries 125 credits a month for the generative work.
Studio tools are the studio agent's own tools, each its own MCP tool under the same name, for one direct step when a whole job is more than you need: generate_image, generate_footage, edit_footage, render_carousel, render_thumbnail, speak, caption_file, web_search, read_page, screenshot_page, x_pulse, news_search and more (tools_list reads the catalog live, with prices). Each takes the write scope and answers {tool, output, media: [{key, url}], chargedUsd}; put a media key in posts_create's mediaKeys to publish it. One with a price spends credit, so confirm first. The full table is at docs/mcp.
Capabilities (capabilities_list) say what ShapelessAI can do, live or planned; capabilities_want records that the user wants a planned one.
Agents are standing schedules (agents_list, agents_save, agents_wake) - recurring content work the server runs on its own.
Brand Memory (brain_tree, brain_read, brain_write) is the account's durable knowledge: voice, positioning, product facts. Edit it when the user corrects how their brand should sound - that fixes every future post, not just one.
Actors (characters_list, characters_voice) are the faces and voices videos use. characters_voice designs a voice from a description or clones one from a sample the user holds the rights to - it spends credit, so confirm first.

Ground rules

Tools whose descriptions say they publish or put content out are live: approving, publishing, waking an agent. Confirm with the user before calling them unless they explicitly asked for that exact action.
Review before approving: read the post's actual text and media via posts_get, don't approve blind.
Generative quality issues (wrong tone, wrong facts) are Brand Memory issues - offer to fix the source, not just the symptom.
Putting a file in the brand library is for reuse; passing its media key on a message is what makes this turn's agent look at it. They are different asks - do the one you were asked for.
The API is rate-limited at 600 requests/hour per key; poll jobs with restraint.
Read posts_create's rules before writing the post, not after a 422: the character limit and the media rules are in platforms_list, and a draft written past the limit is a rewrite.

The full contract - routes, scopes, refusal codes - is at https://shapelessai.com/docs/api (Markdown at https://shapelessai.com/docs/api.md).