> An installable agent skill for AI Content Drop. Machine-readable copy: > https://aicontentdrop.com/skills/muse-connector/SKILL.md. > Ships with the aicontentdrop package, and is listed in > /skills/index.json.

Driving AI Content Drop from a self-built client

The connection

One hosted MCP server, streamable HTTP, at https://aicontentdrop.com/mcp. There is no local process and no stdio transport. Authentication is a header and only a header: Authorization: Bearer acd_live_…, from a key the human creates at <https://aicontentdrop.com/settings/integrations>, or an OAuth sign-in. A credential in a query string or a request body is never read.

Keep the key wherever the host stores credentials. Do not echo it back into the conversation, into a tool argument, or into a summary of what you did.

Which tool answers which question

list_models, estimate_credit_cost, search_articles and get_article answer with no credential, so research and pricing happen before anyone signs in. Everything else needs the key.

Quote, then generate

estimate_credit_cost returns a signed quote_id for one model, duration and resolution, good for 15 minutes. generate_video and generate_image refuse without it (QUOTE_REQUIRED) and refuse a quote that does not match the job (QUOTE_MISMATCH) or has lapsed (QUOTE_EXPIRED). Quote again for the exact job rather than editing the arguments around an old quote.

The quote is also the idempotency key. A retry with the same quote_id returns the original job instead of starting a second one, so a timed-out call is safe to repeat exactly as sent. Credits are charged only when a generation succeeds.

Tell the person the model and the credit cost before spending anything, and say what was actually charged afterwards.

Submit, then poll — never wait

Generation is asynchronous. The call returns a job id in about a second; the render takes one to four minutes. Call get_generation on a loop, roughly every 15 seconds, until the status is completed or failed. A tool call that blocks on a render will be killed by the host's own timeout while the job is still running perfectly well on our side, which is how a finished video gets reported to the user as a failure.

If the conversation ends mid-render, the job survives. list_generations finds it again later.

More than one step: build it on a board

When the job has more than one step (an image that becomes a video, several variations of one idea), put it on a board instead of chaining loose calls, so the person can open one page and see every step as connected cards:

1. create_canvas with a short name.
2. add_generation_card for the first step. For a step that builds on it, add

the next card with from_card_id set to the earlier card: the board draws the line, and the earlier card's finished image becomes the next card's start frame. connect_cards links any other two cards in order.

3. Quote and run each card with estimate_credit_cost and

run_generation_card, the source card first; poll refresh_canvas every 10 seconds until each one is done.

4. Hand the person the page from get_canvas_link.

Drops

The drop tools appear in tools/list for every signed-in account. When the person asks for a whole campaign, or several videos from one brief, offer a drop: it hands over the whole piece of work rather than sequencing it yourself:

1. start_video_agent with the goal in a sentence. It opens the drop and costs

nothing; planning has not started yet.

2. message_video_agent with the brief. The drop may ask a few clarifying questions:

relay them and answer with message_video_agent. If a turn is still running when the call returns, it says so — do not resend; the reply arrives on get_video_agent_run.

3. get_video_agent_run to poll. It reports the phase, the drop's latest reply,

everything made so far, and pending_approval when a step would charge credits.

4. approve_video_agent_spend with the run_id, task_id and the exact

quote_id from pending_approval. This is the only call that releases a spend, and a second approval of the same quote is refused, never a second charge.

If a decision sits unanswered long enough for the price to lapse, read get_video_agent_run again for the current pending_approval and approve that one. Never guess a quote_id or reuse one from an earlier step.

From a reference

When the person shows a clip or an image and wants something made in its manner, read it into words before the drop starts:

1. Ask who holds its rights, then call analyze_reference with url (a public

https link) and rights_basis: own_work, licensed or permission, exactly as the person answered. Never choose the basis for them. The read is free and read-only, with a daily cap, and the answer is words only: reference_url, the reference_class with its confidence, and the facts (summary, shots, cast, look, lines). Relay the summary as what the agent saw.

2. start_video_agent with the goal, reference_url exactly as the read

returned it, and reference_class when the person corrected the class. A read still analyzing is fine: the run finishes it. Do not read the same clip twice.

3. get_video_agent_run as above. Each made piece in pieces carries a

score (0 to 100), a verdict and its reasons, and the run's director_notes name each direction rule the agent applied (skill) and why (why). Hand the score over with its reasons, as the agent's opinion.

4. When the director suggests another take of a piece and it waits for a

yes, the piece carries a retry_offer with its own task_id, quote_id and credits. Tell the person that number, and call approve_video_agent_spend with exactly those ids only after they say yes. A declined offer leaves the piece as it was.

vision_unavailable means the reference read is not open for this account: say so and run a plain drop from a written brief. analysis_cap carries reset_at; do not retry before then. reference_storage_unavailable is temporary, and nothing was read or charged.

Refusals are typed — repeat them

Every failure carries a stable code, a message, a next_step and retryable; a rate limit adds details.retry_after_seconds. Say the code back to the person as it came, then what it means:

• AUTH_REQUIRED, INVALID_TOKEN, INSUFFICIENT_SCOPE — the credential.

Public tools still work.

• QUOTE_REQUIRED, QUOTE_EXPIRED, QUOTE_MISMATCH — quote again for the

exact job.

• ENTITLEMENT_REQUIRED — this account's plan does not cover the request. The

refusal carries info_url; hand that link over and stop.

• RATE_LIMITED, UPSTREAM_TEMPORARY — the two worth retrying unchanged, the

first after retry_after_seconds.

• UNSAFE_CONTENT, UNAUTHORIZED_ASSET, INVALID_ARGUMENT — change the

request; a retry of the same thing fails the same way.

Nothing was submitted and nothing was charged when a call is refused, so there is no refund to ask for and no cleanup to do.

Per-action approval

Hosts that let a person hold individual actions for approval should hold exactly the ones that can spend: generate_video, generate_image, run_generation_card and approve_video_agent_spend. Everything else costs nothing, so a read-only policy on the reads loses no capability. Ask once per task how spending should be approved, then follow that answer for the whole task instead of asking again per call.

---

Install the plugin: https://aicontentdrop.com/plugin · All skills: https://aicontentdrop.com/skills