InVideo AI Agent Two, MCP and ChatGPT: A Practical Playbook

Agent Two is the agent you direct inside InVideo AI in November 2026, and good results from it come down to three choices made before the first prompt: the tier, the approval mode and the context you load. It replaced Agent One in August 2026 and routes each shot to one of 200+ third-party models on an editable timeline. The plan gate comes first. Starter costs $20 a seat a month billed monthly and gets only the Lite tier.^[1] Plus, at $60 a seat a month billed monthly, opens Pro and Ultra. Outside agents can drive the same timeline: ChatGPT, Claude and Gemini connect through the invideo MCP server, while Claude Code, Codex and Cursor use the invideo CLI, both with browser sign-in.

Agent One is retired; here is what Agent Two does

Agent Two arrived in August 2026 as the successor to Agent One, and invideo says it is four times faster and twice as cost efficient. That is the vendor's claim. If an older model page still names Agent One, read it as Agent Two. The agent plans a video from your instructions, writes and storyboards it, generates the missing media and assembles the cut on the timeline as real clips you can reshape or undo by hand. Jobs run in the cloud, so a long generation keeps going after you close the browser tab.

The agent reads your videos, scripts, PDFs, links and rough cuts against the project's Context, which persists for the life of the project. Character sheets, locations and voices are stored and reused, and the agent checks Context before every generation. The original flow, where a pasted script becomes a stock-media video with voiceover and subtitles (see our text-to-video guide), is still reachable through the Older versions button on the prompt page. The default is Agent Two. It runs as a guided conversation through workflows such as Explainer or Script to Video, or in one shot with Autopilot.

Set the tier and the approval mode before you prompt

The tier is one lever on what each generation costs, since Lite is light on credits and Pro and Ultra use more per generation, so use the tiers in sequence. Rough out shots in Lite, then switch to Pro or Ultra only for the takes you plan to keep. Starter cannot do this. It offers Lite only and leaves out Seedance 2.5, so final-quality agent work needs a higher plan. The math stays tight even there: at about 96 credits per 5-second 1080p Seedance 2.5 clip, the 2,000 monthly credits on Plus cover about 20 clips and the 5,000 on Max about 52, before retries, by our arithmetic.

Approval mode is the second lever. Agent Two defaults to always allowing generation, which burns credits fastest when a brief is vague. Switch it to ask before video or before any generation, and an Approve popup shows the prompt, model, duration and aspect ratio before anything runs.^[2] Owners can also cap credits per member, and the agent stops at the cap. invideo's own productions average about three generations per usable shot, and plan credits reset monthly without rolling over (Enterprise lists volume rollover credits), as our pricing breakdown explains.

Brief Agent Two like a crew, not a search box

Load the project before asking for shots. Upload the script, reference images and any rough cut, then have the agent build character sheets and locations into Context, so later generations reuse them. invideo says the agent reviews each generation against your references and regenerates failures, a check you can tune or switch off per shot. Locking which models the agent may use keeps the look and the credit rate predictable. One provider limit matters early: Seedance 2.0 is likely to block reference images showing a clearly visible realistic face.

Playbooks, added in August 2026, turn rules you would otherwise repeat, such as an aspect ratio or a caption style, into standing instructions attached to an agent. Write them once. They hold session after session. Workflows, also from August 2026, are guided jobs; nine shipped at launch, including casting, storyboarding and localization, and you can save your own steps as a markdown file for future projects. The Reframing workflow shows the right habit: it extends frames as images first and waits for approval before generating video, so credits go into approved shots.

When the agent keeps missing one shot, stop prompting and open Notebooks, another August 2026 launch. It lets you work a single generation by hand with close to 200 models, and every generation carries a thread ID you can point the agent at later. Slate then lines up approved takes and applies edit notes you speak. If a project uses cast members, voices or avatars, InVideo AI requires explicit consent from the person in the recording, captured in a separate in-product step that can be withdrawn; celebrities and politicians cannot be cloned, and an avatar of anyone under 18 is forbidden even with parental consent.

Driving invideo from ChatGPT, Claude and Gemini

The invideo MCP server, launched in September 2026, lets a chat app work on the same timeline you see in the invideo Editor. It signs you in through the browser with OAuth, and every call runs as you, inside the workspaces you chose.^[3] The outside agent works on its own branch under a cursor with its name, so the main cut stays unchanged until you accept the change. Ordinary editing through the connection uses no invideo credits; generating images, video or audio does.

Inside ChatGPT

In ChatGPT, invideo is an official app: find it among ChatGPT's apps, sign in, and create an invideo account during sign-in if needed. Whether you can add apps depends on your ChatGPT plan, workspace settings and region, and Business or Enterprise admins may need to allow it. A ChatGPT plan does not include invideo.^[4] ChatGPT's own limits cover the chat, while generation spends credits from your invideo plan. ChatGPT gets the plugin without invideo's skills, so it learns the rules from the server's help guide.

Claude and Gemini

In Claude on the web or desktop, invideo is a custom connector; on Claude Team or Enterprise an Owner adds it first, then each member connects their own invideo account. To let Claude upload attached files, allow m.invideo.io once under Settings, Capabilities, Allowed websites. The Gemini app adds invideo as a custom app, which Google limits to personal accounts of people 18 or over in the US, in English, with Keep Activity on. Whatever the app, the agent sends exact edit values rather than loose commands. Vague requests produce vague timelines. Ask for timecodes, durations and tracks.

Coding agents: Claude Code, Codex and Cursor on the CLI

Terminal agents skip MCP and use the invideo CLI, a single native binary that needs no sudo. In Claude Code you install the invideo plugin from the invideoio/skills marketplace with two terminal commands; on the first video request the agent installs the CLI and opens a browser tab for sign-in. Codex also runs invideo through the CLI. In Cursor, the Marketplace plugin carries only the MCP server, so add the skills with npx skills add invideoio/skills for the CLI route. Both routes work on every plan, like the invideo Editor itself.

On ssh, in a container or on a CI runner, a person approves once with a device code. Keep the routes apart. The CLI and MCP use separate sign-ins and should not be mixed in one session. Every paid call follows one guardrail: a dry run returns the price, a max-credits cap blocks anything costlier, and an idempotency key stops a retry from paying twice. Calls under 1 credit run without asking. Several agents can share a project, each on its own branch, but branches do not merge, and a message to Agent Two runs on your credits.

No outside agent can change the plan, buy credits, cancel, request a refund, share or delete a project, or add people, and none posts or schedules to social platforms; you upload the file yourself. That matters for budgeting, because fees are non-refundable: invideo's Terms allow a refund request only within 24 hours of a first purchase or renewal and only if no credits or premium features were used, and credits are never refunded, as our InVideo AI review details. Start with small edits, since editing and exports are free, while generation, transcripts, paid stock and Agent Two's own work draw on the monthly credits.

The bottom line

Agent Two rewards preparation. Load Context, write Playbooks, keep Lite for drafts and approvals switched on, and it behaves like a crew that remembers the brief; skip those steps and the credits drain into retries. The tier gate makes Starter a testing seat rather than a production one. For people who already work in ChatGPT, Claude or a coding agent, the MCP server and the CLI are the cleaner path: editing through them is free, generation shows a price before it runs (sometimes an estimate that can end higher), and the main cut stays untouched until you accept a branch.