If you searched for the fireship new video march 2026 agents open source tools, you probably want the practical part: which agent tools were shown, what they actually do, and how to run them on your own machine without wiring together a dozen half-finished scripts. This post walks through the tool categories that video covers, then shows the pipeline that turns agent output into a finished video and a published post. We build that pipeline in public on the KevBuildsApps YouTube channel, so you can follow along with real runs instead of slides.
prompt -> coding agent -> files on disk
|
v
skill scripts (Python + ffmpeg)
|
v
rendered short -> upload -> post
^
|
CLI or MCP tool call
What the Fireship March 2026 Video Actually Covers
The short answer: agents that live inside your terminal and edit files, plus the open source glue that lets them touch real things like video timelines, repos and APIs. The video is not a product tour. It is a tour of a workflow, and the workflow has three parts.
First, a coding agent. You give it a task in plain language. It reads your files, writes new ones, runs commands and checks its own output. Second, skills. A skill is a folder of instructions plus scripts the agent can call. Skills are how you stop pasting the same 400 word prompt every morning. Third, local rendering. The heavy work happens on your machine with Python and ffmpeg instead of a cloud render farm with a queue and a bill.
That is the whole shape. Everything else in the video is an example of those three parts doing something you could not do a year ago without a small team.
Why the terminal matters more than the demo
Demos look good because someone already prepared the inputs. The interesting claim is that the agent can run on your machine, with your files, and leave artifacts you can inspect. If a step fails, you look at the file it wrote. That property is what makes an agent useful for production work rather than a party trick.
The Agent Tools Fireship Highlights and Why They Matter
Two categories come up: coding agents and agent harnesses. A coding agent is a general tool that edits code and runs commands. Claude Code is the one we use daily, and it is the one that shows up most in the video. An agent harness is narrower. It wraps a specific job, like running your social accounts, and gives the general agent a set of tools and a memory of what happened last time.
Creator OS ships exactly that second layer for publishing. The hosted MCP server lives at https://mcp.creatoros.ca/mcp and there is nothing to install. Add it as a connector, sign in, pick a workspace, then choose read only or read and write. Read-only connections only see read tools, and the API refuses writes from them, so you can let an agent look at analytics before you let it post.
Add it to Claude Code like this:
claude mcp add --transport http creatoros \
https://mcp.creatoros.ca/mcp \
--header "Authorization: Bearer cos_live_..."Cursor, Windsurf and VS Code take the same entry in mcpServers with type: http and the Authorization header. ChatGPT takes the same URL as a connector in developer mode. The tool names are plain: create_post, update_post, upload_media_from_url, list_posts, get_post_analytics, get_best_time_to_post, list_comments, reply_to_comment, create_automation, check_caption_length, links_create, skool_create_post, blog_create_article, ads_boost_post, skills_list. Destructive ones such as delete post, delete comment, disconnect account and delete ad are marked, so the AI app asks before it removes anything.
Every plan includes API keys, the MCP server, the CLI and the agent skills. Setup steps for each app are at https://www.creatoros.ca/docs/mcp.
Open Source Tools That Pair With Coding Agents
The video’s open source segment is about tools you can read, fork and run offline. Two of ours fit that description and we use them in the same sessions.
Social Agents (https://github.com/kevinbadi/social-agents) is an agent harness that runs your socials on Creator OS. Start it with npm start creatoros social-agents. It interviews you about your brand, then posts, replies to comments and DMs, runs automations and reports back with a local dashboard. You bring your own model: a logged-in Claude Code session, an ANTHROPIC_API_KEY, or any Anthropic-compatible API.
OPEN VIDEO EDIT (https://github.com/kevinbadi/open-edits) is a set of Claude Code skills that turn a raw talking-head clip into an animated short. Captions, logos, faces, GIFs, b-roll cards and counters, rendered locally with Python and ffmpeg. The skills are named things you can say out loud: split-animated-talking-head, longform-animated-talking-head, fireship-style-edit, fireship-style-short, split-animated-short, reverse-engineer.
There is a walkthrough of both on the launch video. We also wrote about the editing side in Fireship style video editing with AI and about the harness in Social Agents open source.
A worked example: one clip, one short
Record a two minute talking head. Open Claude Code in the repo. Ask for the split-animated-short skill on that file. The agent runs the split, picks the beats, adds captions and cards, calls ffmpeg and leaves an MP4 in your output folder. You watch it, ask for a change, and it re-renders. No timeline, no export dialog.
How This Fits the 2026 Agent Stack for Creators
Here is the stack as it stands. A coding agent at the top. Skills in the middle for repeatable jobs. A publishing layer at the bottom that knows platform rules. The middle layer is where most people stop, because skills are easy to write and hard to maintain. The creatoros sync command updates your skills without overwriting files you edited. Updates land as .new files so you can diff and merge.
Run npx @creatoros/cli@latest init and you get 14 Claude Code skills installed at once: post-everywhere, post-shortform, post-longform, post-threads, schedule-posts, respond-to-comments, respond-to-dms, automations, analytics, kairos-dashboard, ads, meta-campaign-builder, skool, blog.
On the publishing side, Creator OS covers Instagram, TikTok, YouTube, X, LinkedIn, Facebook and Threads, plus Skool communities, a WordPress blog and paid ads as an add-on. The CLI is the part that fits an agent stack best, because every command returns text an agent can read.
creatoros posts:create --text "clip notes in caption" \
--platforms instagram,tiktok,youtube \
--media ./out/short.mp4 \
--scheduledAt 2026-03-18T14:00:00Z --timezone America/TorontoThen check the result before you trust it:
creatoros posts:list --status scheduled
creatoros analytics:best-time --platform tiktokTurning Agent Output Into Finished Video
This is the part the video underplays and the part that decides whether the pipeline survives week two. Renders fail. Captions land on the wrong word. A platform rejects a file. You need a retry story.
Keep raw footage and intermediate files on disk, not in a chat. When a render breaks, the agent can read the error, fix the script and run again without you re-recording anything. For published posts, creatoros posts:retry <postId> resubmits a failed post. creatoros posts:edit <postId> --platform youtube --text "new title" fixes text on a live post where the network allows it.
If you upload one file to many platforms, upload it once with creatoros media:upload ./out/short.mp4 and reuse the returned med_ id everywhere. Networks named in a post that are not connected come back in missing_platforms, and the rest still go out. That one behavior saves a lot of confusing half-failures.
Two related deep dives if you want the editing detail: auto captions for Reels with an open source editor and the open source Claude video editor. For a longer pipeline view, see running generation locally.
Where Fireship Stops and Creator OS Starts
A video about agents and open source tools ends at the terminal. It shows the render. It rarely shows what happens after, which is the part that eats your afternoon: captions that are too long for X, a TikTok that needs consent flags, a Reel that needs a cover frame, a comment thread you have not touched in three days.
Creator OS ships a hosted MCP server and a CLI so an AI agent can run your accounts, with those rules applied for you. A dashboard tool gives you a calendar and a composer. Creator OS gives an agent the same reach you have in the web app, and the rules live in the API rather than in your memory.
A few concrete examples. Post to X, Instagram and TikTok with one create_post call and it handles the per platform differences. Pull get_best_time_to_post before you schedule. Read list_comments, then reply with reply_to_comment on Instagram, Facebook, X, Threads, YouTube and LinkedIn. TikTok replies are not supported, and we say so rather than let you find out. DMs work on Instagram, Facebook and X, and all of them land in one inbox.
For comments, the CLI is quick:
creatoros inbox:comments --platform instagram --since 2026-03-11
creatoros inbox:reply post_abc --accountId acc_123 \
--message "thanks, link is in the caption" --commentId cmt_789Short links get click tracking through links_create and links_stats. Analytics cover per post numbers, daily account metrics, follower stats and best time to post, across every connected platform.
Common Mistakes When Adopting Agent Tools
Four show up again and again in our own runs.
- Giving write access on day one. Connect read only first. Read-only connections only see read tools and the API refuses writes. Watch what the agent pulls, then open writes.
- Letting one key see everything. API keys are pinned to one workspace, which is one set of socials. An agent holding one workspace’s key cannot post to another. Use that boundary on purpose.
- Ignoring the platform rules until publish fails. Run
creatoros validate:post-length --text "..."before you schedule. Threads caps posts at 500 characters and 250 posts per 24 hours. YouTube titles max at 100 characters. X video caps at about 2 minutes 20 seconds on standard accounts. - Trusting a render you did not watch. Watch every output before it goes out. The agent will not catch a bad take.
One more: do not build your whole system on a model you cannot swap. Creator OS works with any MCP capable AI app and does not ship its own model. Current Claude models are Claude Opus 5.5, Claude Sonnet 5.5, Claude Haiku 4.5 and Claude Fable 5.1. Your stack should survive a model change.
Get Started With Creator OS Today
The Creator plan is $14.99/month or $59.99/year and covers 11 platforms plus Skool, one account on each. API keys, the MCP server, the CLI and the agent skills are included in every plan. Agency plans step up to 5 sets of socials for $49.99/month or $399/year, and 10 sets for $99/month or $599/year. If you run several YouTube channels, network plans are $19.99/month for 5 channels or $29.99/month for 10. The ads add-on is $19.99/month on any plan, with no percentage of ad spend, and ad spend is billed by each network to your own ad account.
Start with two commands:
npx @creatoros/cli@latest init
creatoros auth:checkThen read the docs at https://www.creatoros.ca/docs for the REST base URL and request shapes, and https://www.creatoros.ca/docs/cli for the full command list. If you want the guided version, watch the walkthroughs on the KevBuildsApps YouTube channel and sign up at https://www.creatoros.ca/sign-up.
Get started for $0
Free trial for new subscribers. Cancel anytime.