presenter

presenter

Bisque's open-source Agent Skills for building narrated presentations youpublish to a shareable watch URL on bisque.today.Narration is synthesized on your own machine — free and unlimited; nothingis billed per word.

The collection ships three skills. present is the main one: you describe apresentation, your agent authors it, your machine narrates it, and Bisque hostsit. It works in Claude Code, Gemini CLI, and Codex CLI, on macOS,Linux, and Windows. watch reads a published presentation — anyone's — soyour agent can summarize it or answer questions about it. video turns anypublished presentation into an MP4 you can upload; it runs on macOS.

How it works

1. Ask your agent for a presentation — a briefing on your repo, a walkthroughof a pull request, a report from whatever it just did.
2. The agent writes the slides and the narration script, and your machineturns the script into speech.
3. You get a watch URL — slides with synchronized narration, playable in anybrowser, on your own channel.

Example

Ask, in any of the three CLIs:

Give me a two-minute briefing on what changed in this pull request.

The agent writes an index.html — one <section> per slide, the narrationfor each slide in an <aside class="notes"> — synthesizes the speech on yourmachine, and publishes. It prints a watch URL on your channel; open it and theslides play with synchronized narration in any browser.

Change a slide and ask it to publish again, and only that slide isre-synthesized — an HTML-only edit costs no synthesis at all. The presentationis a living document, not a one-shot render.

Reading one back

A presentation is a narrated document, so your agent can read one instead ofwatching it. Paste a link and ask:

What does this presentation say about pricing?

The watch skill fetches the full narration transcript and the backgroundcontext.md the author shipped with it, then answers from that material andcites the slides it used. It works on anyone's public presentation, and itneeds no account — sign in and your own private presentations open too.

Turning it into a video

Ask for an MP4 and the video skill takes over from the watch URL:

Render that as a vertical video I can post.

It downloads the presentation and the narration you already published, drawsevery frame with the same player the watch page uses, and writes an MP4 —landscape, square, or vertical for Shorts, Reels and TikTok. Nothing isre-narrated, so the video sounds exactly like the watch URL.

Rendering takes about as long as the presentation runs, and it needs macOS.

Somewhere without a terminal

The three skills above narrate on the machine they run on, which needs aterminal and a few gigabytes of disk. Agents that run in a hosted sandbox haveneither, so this repo ships a second plugin — Bisque, undercowork/ — that publishes the same presentations through Bisque'sconnector, with the recording done on Bisque's side. Same format, same watchURL, nothing to install.

From your own code

The skills call a public REST API, and so can you. @bisquecloud/sdk is theTypeScript client for it — publish from a script, a CI job, or a server, andread any shared presentation back as text:

npm install @bisquecloud/sdk

import { Bisque } from "@bisquecloud/sdk";

const bisque = new Bisque({ apiKey: process.env.BISQUE_API_KEY });
const created = await bisque.presentations.create({ indexHtml });
const { webUrl } = await bisque.presentations.waitUntilReady(
  created.presentationId,
);

It lives in sdk/; the API it wraps is described atbisque.today/openapi.json.

Install

Both plugins live in one marketplace. In Claude Code:

/plugin marketplace add bisque-cloud/presenter
/plugin install presenter@bisque-cloud

In Claude Cowork, open Customize → Plugins → Add marketplace, enterbisque-cloud/presenter, and install Bisque. It connects to your accountthe first time you use it.

Just the skills

Add them with the skills CLI:

npx skills add bisque-cloud/presenter

It copies them into your agent's skills directory — ~/.claude/skills/ forClaude Code, ~/.agents/skills/ for Codex and Gemini. Then ask for apresentation, or invoke it directly with /present.

In VS Code and Cursor

VS Code and Cursor connect to Bisque over MCP rather than installing theskills. Ask either editor for a presentation and you get back the same watchURL. One click adds the server:

Each badge adds a server called bisque pointing athttps://bisque.cloud/presentations/mcp, and the editor walks you throughsign-in the first time you use it. Bisque narrates on its side here, so thereis no speech model to download. Any other MCP client takes the same URL overstreamable HTTP.

Manual install

Each skill is a plain directory; installing one is putting it where your agentlooks. Clone the collection and copy the skills you want:

git clone https://github.com/bisque-cloud/presenter
cp -r presenter/skills/present ~/.claude/skills/present   # Claude Code
cp -r presenter/skills/present ~/.agents/skills/present   # Codex + Gemini
cp -r presenter/skills/watch   ~/.claude/skills/watch     # optional
cp -r presenter/skills/video   ~/.claude/skills/video     # optional, macOS

git clone https://github.com/bisque-cloud/presenter
Copy-Item -Recurse presenter\skills\present $env:USERPROFILE\.claude\skills\present

Invoke with /present (Claude Code), $present (Codex), or let it activate ondescription match. Gemini CLI also reads ~/.agents/skills/, so the Codex copycovers it too. Codex CLI doesn't include Node — install Node 18+ (or bun)alongside it.

Requirements

• Node 18+ or bun.

• bisque-voice — the local narration engine(~30 MB). The agent offers to install it on first use.

• A speech model, downloaded once. There is deliberately no default: theagent shows you the catalogue and you pick. These are large — pick with yourdisk and memory in mind.

  Voice	Download	Peak memory, model only
  Small and fast — nine languages, exact word timings	337 MB	not measured
  Nine ready-made voices, ten languages	1.26 GB + 1.05 GB aligner	not measured
  Clone your own voice, ten languages	1.28 GB + 1.05 GB aligner	~4.9 GB
  Nine ready-made voices, larger model	2.33 GB + 1.05 GB aligner	~5.7 GB
  Clone your own voice, larger model	2.37 GB + 1.05 GB aligner	~5.7 GB

  The four larger voices narrate with measured word timings, so each installs aforced aligner alongside itself — that is the second number, and it is oneshared pack that downloads once however many of them you use. It loads on topof the model while it measures, so peak memory in practice is above thefigure shown. The agent names each download and its size before starting it.

  The small fast voice needs an extra pronunciation pack per non-Englishlanguage — under a megabyte for most, 49 MB for Japanese — downloaded thefirst time you narrate in that language.

• A Bisque account, for publishing. During its first machine check theagent walks you through browser sign-in, then asks you to claim a publichandle — presentations are published under it. Both happen before it writesany slides. Your first handle can never be renamed or released, so pick oneyou are happy to keep; additional handles can be added at bisque.cloud.

• A Mac, for video. It downloads bisque-video (~15 MB) on first use, andan ffmpeg (~12 MB) if the machine has none — a Homebrew ffmpeg already thereis used as-is. present needs none of this.

Limitations

• First use downloads a speech model. See the size table above. The firstsynthesis is not instant.
• The larger voices need their aligner to time cues precisely. They shipapproximate word timings of their own, so they install a forced aligner withthe model to measure instead of infer. --align none tells the synthesizerto fall back to the engine's own timings, but the aligner still downloadswith the engine — it is part of what that voice is. Expect real drift if youskip it: these models emit frames rather than word boundaries, and the clipcan open with untrimmed silence the aligner would have cut. The small fastvoice produces exact timings natively and needs no aligner at all. Slidesthemselves never desync — each has its own audio — but narration inside aslide can start late without alignment.
• Narration is local; hosting is not. Synthesis runs on your machine forfree, but publishing goes to Bisque's servers — there is no self-hostedwatch page.
• Codex CLI's default sandbox restricts network. Installing, signing in,and publishing will surface approval prompts under default sandboxsettings.
• Video rendering is macOS-only, and it is not instant. A render takes aboutas long as the presentation runs — every frame is drawn and snapshotted. OnLinux or Windows there is no MP4, but the watch URL plays in any browser.
• A video needs the presentation to be public or unlisted. video fetchesit by URL like any viewer does, so a private presentation cannot be renderedfrom another machine.
• Intel Macs cannot run the small fast voice, or any aligner. Both arebuilt on a runtime that dropped Intel macOS. The four larger voices — bothready-made-speaker models and both cloning models — all work, so an IntelMac still gets nine voices without recording anything; its cue timings juststay approximate.

Troubleshooting

Run node present.mjs doctor. It checks the whole chain — binary, installedvoices, credentials, account, handle — and synthesizes a test word to proveaudio actually works, which catches the common failures before a publish does:no voice installed, no handle claimed, or a voice that runs but producessilence (on macOS, retry with --device cpu).

For video, run ~/.bisque/bin/bisque-video doctor — it reports the ffmpeg it will encodewith, where the player bundle is, and whether both are usable.

Beyond that, a failed publish is usually no Node on the machine (Codex CLI),no network approval (Codex sandbox), or not signed in. For anything else,open an issue with what youran and what happened — the sentence that misbehaved, if narration soundedwrong.

License

MIT