MCP server

The hosted MCP server is https://shapelessai.com/mcp. Add that URL to any host that speaks remote MCP and it opens a Shapeless tab to sign in and allow. OAuth, no API key, revocable by the human. A client that cannot open a browser (the Agent SDK, a CI job, a self-hosted agent) sends an API key from Settings → API keys as Authorization: Bearer slk_... to the same URL instead.

It is the same operations as the API and the CLI, discovered as tools - so an agent host learns the surface instead of reading this page.

Add it#

Any other host that speaks remote MCP takes the same URL: https://shapelessai.com/mcp.

Every step above was read from the vendor's own documentation; the dated sources are on shapelessai.com/connect.

The handshake, for an agent adding it itself#

• Transport: streamable HTTP at https://shapelessai.com/mcp
• Auth: OAuth 2.1, dynamic client registration, PKCE
• Protected-resource metadata: https://shapelessai.com/.well-known/oauth-protected-resource/mcp
• Scopes: read, write, publish, offline_access
• Instructions: the initialize result carries a short brief for the model (what Shapeless is for, where to start, the approval rule, that the live docs are at /llms-full.txt and the playbook at /skill.md), so a host knows when to reach for these tools before it has read a page. me answers next: the moves worth making on this account now.

An unauthenticated call answers 401 with WWW-Authenticate: Bearer resource_metadata="..." pointing at that document. The server is stateless - one server per request, no session id - so nothing breaks when the next request lands on another instance.

The tools#

Every tool's description names the scope its credential needs. Tools that put content out say so in capitals and carry a destructive annotation, so a host asks the human first. Read-only tools are annotated read-only and run freely.

Studio tools#

Every tool our own studio agent holds is also its own MCP tool, under the same name, with the engine's JSON Schema as its input: images, AI video, video edits, carousels, speech, subtitles, web pages, X and news research. Each is POST /api/tools/{name} in tool clothing and takes the write scope. A call holds budgetUsd of credit (default 2, max 10) while it runs and settles at what it really cost; one with a price spends credit and is annotated destructive, so the host asks first. Every one answers {tool, output, media: [{key, url}], chargedUsd}, where media lists each file it made; put a key in posts_create's mediaKeys to publish it.

The prices are a default call's engine cost before the plan's margin; premium image quality and longer or voiced footage cost more. tools_list reads the same catalog live.

posts_create is POST /api/posts in tool clothing: connectionId, text (the route's postText), mediaKeys, title (the route's documentTitle; YouTube requires one), scheduledAt or queue: true, settings, firstComment, audio (an Instagram Reel's sound, from instagram_audio_search). Read platforms_list first - it carries the limits and the settingsSchema to fill. On the Free plan the eleventh post for one UTC day answers 402 free_daily_cap, which names the day and when it resets. See Posts.

References#

A reference set is the people an account works from: creator, competitor, or peer (creators like me). Add someone with references_add ({handle, platform} or {url}); adding the same person twice answers the member already there. Then start a job from them: jobs_create {goal, referenceId: <member id>, mode: "clone" | "format" | "inspiration"}. The post it makes carries inspiration {kind: "reference-post", memberId, mode}, so its result can be read against the account's own baseline and against posts made from Brand Memory alone.

references_suggest returns {suggestionId, platform, handle, kind, reasons: [{code, evidence}], rank}; show the reasons to the human, then accept with references_add {suggestionId} or dismiss with references_remove {suggestionId}. The list is empty, with a note, until the suggestion job has run for the account.

Picks are private to the account: nobody is told they were picked, and references_browse shows other workspaces only as a count of how many picked a handle, and only at 3 or more. With platform and an exact handle it also answers what those workspaces picked beside it. Browsing is counted per UTC day: Free 5, Spark 50, Growth 500, Studio unlimited. The answer says what is left (remaining); past the limit it is 402 reference_browse_limit with the reset time and the upgrade links.

Two tools exist only on the local stdio server, because they read your disk: assets_upload, brain_import, and the files argument on a job message. Run it with shapeless mcp - see CLI.

Eight Shapeless Directory tools need no key and no account: directory_requirements, directory_preview, directory_submit, directory_verify, directory_stats, directory_search, directory_get and directory_alternatives. They are served by shapeless mcp and alone on the keyless endpoint https://shapelessai.com/mcp/ai-tools, never on this OAuth server. See Shapeless Directory.

Work passes both ways#

• Every job result carries url - https://shapelessai.com/studio/c/<id> - the conversation the human opens. jobs_create, jobs_list, jobs_get and agents_wake all decorate.
• jobs_brief is the token-compact read before replying: the last 30 messages clipped, reasoning and tool-activity dropped, an artifact inventory, post counts per queue status. Deterministic, no model in the loop.
• jobs_list pages 200 at a time with a nextCursor the tool takes back as before.
• jobs_tail replays from a cursor and follows live, returning when the run ends, 60 seconds pass or 200 events arrive. A run with no stream answers {live: false} rather than erroring.

The panel: show the person things, and ask them#

On the Ask page the chat has a panel beside it: a read-only document of tabs the person looks at while the conversation goes on. The browser draws it; the chat's own agent and you, from a terminal, write it, and both see the same panel because it is stored with the conversation.

• No id needed. Leave conversation_id out and the panel tools use the conversation the person opened last in the browser, the one they are most likely looking at.
• panel_open takes kind, title and the kind's fields flat: calendar (ids, highlight, mode), search (query, hits, groupBy, orderBy), prompter (sections), media (items), node (node), performance, agent, people, brain, plan, choices, and custom (html, drawn in a sandboxed frame), and flow (nodes, edges, estimateUsd, approve: a plan as a graph, one node per choice with its cost and what it was built from). The same id replaces a tab in place.
• panel_ask is the one way the person answers you there: two or more options, one press. With wait_seconds it returns picked: {optionId, label, at}; null means not yet, so keep waiting with panel_watch and after_rev.
• panel_watch with after_rev and wait_seconds holds up to 25 seconds and returns the moment the panel changes. Its text is one plain line per tab, its picks every answer.

The same panel is plain HTTP: GET /api/conversations/{id}/panel (?after=<rev>&wait=<s>), POST /api/conversations/{id}/panel with {ops: [...]} (an edit op there is the person's hand edit of one flow node: {id, node, patch {text?, sub?, media?}}), and GET /api/conversations/current/panel. From a shell: shapeless panel show | open | update | close | ask | watch.

There is one prompt, continue (argument: id), which a host like Claude Code surfaces as a slash command: it loads that conversation's brief and tells the agent to reply into the same thread with jobs_continue.

Claude Code plugin#

The public repo is also a plugin marketplace. The plugin wires up the hosted server and ships a skill that teaches Claude the ropes:

/plugin marketplace add FirstClassTree/shapelessai
/plugin install shapeless@shapeless

Then run /mcp, pick shapeless and choose Authenticate.

Other servers#

We keep a measured survey of every MCP server that posts to social networks, ours included, at shapelessai.com/best-social-media-mcp-servers: endpoint, auth, which tools publish, networks, and the free tier's honest limit, each read from the vendor's own page.

Questions#

Is there a free social media MCP server?#

Yes, this one. https://shapelessai.com/mcp is on the Free plan: OAuth, no API key, 10 posts a day across nine networks, no card. The same tools run on every paid plan.

Can Claude post to social media?#

Through a server that exposes a publishing tool, yes. Add this server as a custom connector, connect an account under Accounts, and ask. posts_create publishes or schedules; it is annotated destructive, so Claude asks you before it runs.

Does it work in ChatGPT, Cursor and Codex too?#

Yes. In ChatGPT, Shapeless is a published plugin: open the Shapeless plugin while signed in, add it and sign in. Any other MCP host that speaks streamable HTTP with OAuth 2.1 can add the server by its URL; the host table above lists the ones we have verified step by step. Claude Code also has a plugin, and the shapelessai CLI runs the same tools over stdio for hosts that cannot reach a hosted server.

What happens when the free plan's daily cap is hit?#

posts_create answers with the free_daily_cap error, the limit and the time it resets, so an agent can wait rather than retry blindly. Nothing is dropped silently. A post that failed still holds its slot on that day, so retry it with posts_publish on the same id rather than calling posts_create again.