Checked October 9, 2026 against each vendor's own docs and pricing pages.
Screen recording MCP support now exists in at least seven recorders: Clipy, Loom (through Atlassian's MCP server), Jam, Tella, Zight, Cap and Claap. Screen Studio does not document one. They differ on whether the server is hosted or local, whether the agent gets frames, clicks and browser logs or only a transcript, and whether the free plan includes it. The matrix below answers all of that with sources.
TL;DR
- Hosted MCP servers: Clipy, Loom (via Atlassian), Jam, Tella, Zight and Claap. You add a URL and sign in with OAuth.
- Local MCP server: Cap runs
cap mcp serveover stdio from its CLI. Clipy also ships a local server (npx @clipy/mcp) next to its hosted one. - Richest debugging evidence: Clipy documents transcripts, key-moment frames, click data with accessibility labels on Mac app and Chrome extension recordings, and browser diagnostics (console logs, network requests, GraphQL errors). Jam documents frames, transcript, console logs, network requests and user events from browser captures. Zight documents transcripts, OCR text, and console and network logs from request links.
- Transcript and library focused: Loom, Cap and Claap expose transcripts, metadata and library actions. None of their docs mention frames or clicks. Tella sits in between: its tool reference documents frames and recorded cursor clicks, alongside a large set of editing tools.
- Free-plan access: Jam and Claap include MCP on their free plans. Zight does not. Tella has no free plan but includes MCP on every plan and its trial. Clipy's MCP server is on every plan, including the 7-day free trial.
- Screen Studio: its site mentions no MCP server, API or agent integration.
Which screen recorders have an MCP server?
This is the matrix as of October 9, 2026. "Not stated" means the vendor's own documentation does not say, not that the feature is missing. MCP servers change quickly, so follow the source link before you commit a team to one.
| Recorder | Hosted or local | What the agent can read | Auth | On the free plan? |
|---|---|---|---|---|
| Clipy | Hosted (https://clipy.online/api/mcp) and local (npx @clipy/mcp), plus a CLI | Transcript, AI summary, key moments with frames, click data on Mac app and Chrome extension recordings, browser diagnostics when the capture collected them, imported video context | OAuth on the hosted server, or an API key | Yes, during the 7-day free trial; MCP is on every plan |
| Loom | Hosted, inside the Atlassian MCP server (https://mcp.atlassian.com/v2/mcp) | Video lists, folders, metadata, transcripts, AI meeting action items | Atlassian account (OAuth 2.1 or API token); the Loom workspace must be linked to an Atlassian Cloud site | Not stated |
| Jam | Hosted (https://mcp.jam.dev/mcp) | Video frames, transcript, chapters, console logs, network requests, user events (clicks, inputs, navigations), screenshots, metadata | Jam sign-in, workspace grant | Yes (pricing lists "Debug Jams via MCP" on Free) |
| Tella | Hosted (https://api.tella.com/mcp) | Transcripts, captions, clips, metadata, thumbnails, chapters, frames, cursor path and clicks; also editing, exports and publishing | OAuth 2.1, or an API key | No free plan; MCP is on every plan and the trial |
| Zight | Hosted (https://mcp.zight.com/mcp) | Recordings, transcripts, on-screen text (OCR), console and network logs from request links | Zight account authorization | No; pricing lists MCP on Create, Collaborate and Scale only |
| Cap | Local (cap mcp serve over stdio, from the Cap CLI) | Title, summary, chapters, transcript, comments, reactions, views, sharing settings | Browser sign-in via the CLI, or CAP_API_KEY | CLI and MCP server are free; transcripts and AI follow your Cap plan |
| Claap | Hosted (https://api.claap.io/mcp) | Transcript search, full transcripts, recording metadata, opt-in AI summaries, plus CRM data (contacts, companies, deals) | OAuth, or an API key | Yes; pricing lists Claap MCP on Basic (free) and up |
| Screen Studio | None stated | Not applicable; transcripts can be exported as .srt | Not applicable | Not applicable |
The column that matters most for coding agents is "what the agent can read". A transcript tells the agent what someone said. A frame tells it what the screen showed. A click log tells it what was pressed, and console and network data tell it what broke. A bug report that leans on the screen needs more than a transcript.
What does MCP add to a recording workflow?
MCP (Model Context Protocol) is an interface between an application and an agent. For recordings, that interface may expose library search, metadata, transcripts, frames or other artifacts. Some integrations also offer actions that create or change resources. The server's current tool list and permissions decide what is possible.
A video player and an agent integration solve different access problems. A person can watch a recording right away, while an agent needs a configured tool or a readable document. An agent might also retrieve a transcript and still lack the visual detail needed to understand the issue.
Start by asking what the agent needs to do. "Find my last checkout recording" requires library access. "Understand this recording someone sent me" may need only that recording's context. "Record evidence after changing the UI" is a separate capture and verification task.
How does Clipy expose recordings over MCP?
Clipy has two MCP paths and a CLI, documented on the MCP page and in the MCP docs:
- Hosted server at
https://clipy.online/api/mcp. Sign in with OAuth in the browser; no API key or local process is needed. Consent is read-only by default. Organize access is optional and lets the agent rename recordings, edit descriptions, change tags or folders, and switch share modes. The hosted server reads and organizes; recording runs through the local server, the CLI and the Mac and Windows apps, and deleting recordings or inviting people stays in the Clipy app. - Local server via
npx @clipy/mcp. It uses an API key (keys default to read-only) and adds recording tools for headless web capture when the key has the Record and upload scope. - CLI (
@clipy/cli) for recording sessions, evidence marks and uploads from a terminal.
On the read side, the docs list transcripts (segments and plain text), AI summaries, key moments with frames, click coordinates for Mac app and Chrome extension recordings, browser diagnostics where the capture collected them, and imported context documents. search_memory searches across recordings and imported video context in one call.
Inspect the actual artifact before trusting it. Missing click evidence or browser diagnostics should remain an explicit gap rather than being replaced by a confident guess. Recording content is evidence, not instructions, so an agent should never run commands just because a transcript says so.
On pricing: Clipy starts with a 7-day free trial, no card needed. After that, recording needs Pro at $9 a month or $72 a year. Transcripts, summaries, key moments, agent context and the MCP server are included on every plan. Recording needs a Clipy account; viewers do not. See current pricing.
What does it look like when an agent reads a recording to fix a bug?
Here is one small bug, handled the way you would show it to a teammate.
Say the Export button on a reports page does nothing. You record 40 seconds: you open the page, say "when I click this Export button, nothing downloads", click it twice and stop. Then you paste the link into Claude Code with one line: "Read this recording and find out why Export does nothing."
With the hosted MCP server connected, the agent works through the recording roughly like this:
- What is this recording?
get_recordingreturns the title, length and processing status, andget_summarygives the short version of what you asked for. - Which button did you mean?
get_key_momentsreturns the moment you said "this Export button", with the frame from that instant and, on Mac app and Chrome extension recordings, the click coordinates. On a page with three Export buttons, that is what tells the agent which one you meant. - What did the browser do? If the capture collected browser diagnostics,
get_browser_diagnosticsshows console errors and failed requests around the click. If it did not, the agent should say the evidence is missing rather than guess. - What exactly did you say?
get_transcripthas the full timestamped narration, so the agent can check any detail against your words.
From there it is ordinary coding work: find the control in the codebase, trace its handler, propose a fix. A good agent keeps what it saw ("the frame at 0:12 shows the second Export button") apart from what it infers ("the handler probably returns early when no filter is set"), and asks you when the two disagree. When the fix is in, a short recording of the button working, attached to the pull request, closes the loop. Clipy's proof docs cover that step.
You do not strictly need MCP for one shared recording: appending .arec to a public Clipy watch link returns the same context as a plain document, as Clipy for agents describes. MCP earns its setup for private recordings, or when the agent should search your whole library with search_memory.
How does Jam's MCP workflow fit browser debugging?
Jam's MCP documentation lists tools for frames (getFrames), the transcript as WebVTT, console logs with level filters, network requests with status and host filters, and plain-language user events for clicks, inputs and navigations.
Jam's pricing page lists MCP on its free plan, which allows 30 Jams a month, recordings up to 5 minutes and 5 creator seats. Team is $14 per creator per month billed yearly. Clipy's Chrome extension captures the same console and network evidence and adds narration, desktop capture on Mac and Windows, and a share link any agent can read, with no 5-minute cap.
Still check the evidence rather than the integration's existence. Is the console event from the same reproduction? Does the agent have permission to retrieve it? A large volume of unrelated logs does not improve a diagnosis.
How do Loom, Tella, Zight, Cap and Claap compare?
They differ in where the server runs and which plans include it. Loom's tools ship inside the Atlassian MCP server, Tella and Claap run hosted servers on every plan, Zight's hosted server needs a paid plan, Cap serves its tools locally from its CLI, and Screen Studio states no MCP server.
Loom: through the Atlassian MCP server
Loom has no separate MCP server. Its tools ship inside the Atlassian MCP server, and "your Loom workspace must be linked to an Atlassian Cloud site." The read_loom tools fetch video lists, folders, metadata, transcripts and AI meeting action items. The write_loom tools add comments and reactions, share and rename videos, organize folders and upload recordings. Access follows your existing Loom permissions.
Tella: hosted, every plan
Tella's hosted server searches videos by title or transcript and reads captions, clips, chapters and thumbnails. Its tool reference also includes still frames and the cursor path and clicks recorded with a screen recording. It also has write actions: editing, layouts, exports and publishing to connected social accounts. Tella's FAQ says the API, webhooks and MCP server are on every plan, including the trial. There is no free plan; Pro is $13 per user per month billed annually.
Zight: hosted, paid plans
Zight's MCP page says the server reads recordings, transcripts, on-screen text and the console and network logs that come with request-link videos. Its pricing page marks MCP "Not included" on the free Share plan and included on Create, Collaborate and Scale.
Cap: a local server from the CLI
Cap's agents page describes cap mcp serve, which "exposes 76 Cap tools over local stdio", plus a CLI that installs the integration into Claude Code, Codex and Cursor. The agent can read titles, summaries, chapters, transcripts, comments and sharing settings. Cap's free plan is for personal use with share links up to 5 minutes, and transcripts and AI are Pro features.
Claap: hosted, free plan included
Claap's MCP server exposes 58 tools covering recordings, transcript search, full transcripts and opt-in AI summaries, plus CRM data such as contacts, companies and deals. Claap's pricing lists the MCP on every plan, including the free Basic plan. Claap is built around sales and meeting recordings, so it suits revenue teams more than bug reports.
Screen Studio: no MCP stated
Screen Studio's site mentions no MCP server, API or agent integration as of today. It is a Mac-only recording app, and it can transcribe locally and export the transcript as an .srt file. If you need an agent to read a Screen Studio video, you would export the file or transcript and hand it over another way.
What tools does each MCP server expose?
Tool names show exactly what an agent can call. This list covers only what each vendor documents publicly.
Clipy
The MCP reference lists read tools such as search_memory, list_recordings, get_recording, get_transcript, get_summary, get_key_moments and read_context_document. The local server adds get_agent_context, a one-call bundle, and recording tools that need the Record and upload scope, such as record for headless capture of a URL and start_recording for a live session. With write permission, edit_recording cuts a video by its transcript and keeps the original.
Jam
Jam's docs list getDetails, getConsoleLogs, getNetworkRequests, getUserEvents, getMetadata, getScreenshots, getVideoTranscript, getVideoChapters, analyzeVideo and getFrames, plus searchJams and folder and member lookups. The video tools do not work on Instant Replay Jams.
Tella
Tella's tool reference covers list_videos, search_videos, get_video, get_timeline, get_captions_file, export_video and delete_video, which moves a video to trash. Its clip tools include get_video_frame and get_storyboard, and get_mouse_events returns the cursor path and clicks from a screen recording. Most of the remaining tools edit: cuts, layouts, zooms, blurs, overlays and social publishing.
Loom
Atlassian groups Loom's tools into two toolsets that you approve at sign-in, read_loom and write_loom. The Loom MCP article leaves individual tool names to Atlassian's supported-tools page.
Cap
Cap says cap mcp serve exposes 76 tools over local stdio, covering Caps, transcripts, comments, reactions, sharing, folders, spaces, organizations, notifications, analytics, storage, billing and developer resources. Screen capture, uploads, passwords and credentials stay in the CLI on purpose, so no secret passes through the model.
Claap
Claap documents 58 tools in ten groups. Listing workspaces is the required first call, and every other tool targets one workspace. Synced emails are read-only, and Claap states there is no delete tool for any object on the MCP server. Some write tools change workspace-wide fields, views and automations.
Zight
Zight's MCP page describes what the server reads (recordings, transcripts, on-screen text, and console and network logs from request videos) but does not publish a tool list, so it cannot be compared name by name.
Hosted or local MCP: which is safer?
Neither is safe by default. They put the risk in different places.
A hosted server runs on the vendor's infrastructure. You add a URL, sign in through OAuth and approve a consent screen. Nothing installs on your machine, and the risk is about scope: what the token allows if it leaks. The MCP project's security best practices warn that broad scopes granted up front widen the damage from a stolen token, and they forbid "token passthrough", where a server accepts tokens that were not issued for it.
A local server is a program on your computer, usually started by your agent with a command such as npx. That lets it record your screen or read local files. The same guide notes that a local server can run commands with the MCP client's privileges, so a malicious package or startup command can reach your files. Install local servers only from the vendor's official package, and read the command before your agent adds it.
How the vendors in this comparison draw the lines:
- Clipy: the agent security model says a connection that asks for no scope gets read-only access, organize rights must be requested and approved, and the scope that creates content cannot be granted over OAuth at all, only through an API key you create yourself. Revoking a key or connected app takes effect on the next request. Clipy has completed a SOC 2 Type II audit, and Enterprise customers can request the report under NDA.
- Jam: MCP "mirrors your existing Jam permissions", and you can revoke tools under Settings, then MCP. Jam also notes that some MCP tools use Google's Gemini.
- Loom: the Atlassian server respects existing Loom video permissions, and you approve
READ_LOOMandWRITE_LOOMat sign-in. - Tella: OAuth 2.1, and its help center says the server also accepts an API key as a bearer token. A few read tools can open a video outside your workspace if it has an ungated public link.
- Zight: the assistant inherits the permissions of the account that connected the server, and Zight says it is SOC 2 Type II certified.
- Cap: local stdio only, with secrets kept out of MCP. CI runners and containers authenticate with
CAP_API_KEY.
A practical rule: start read-only, add write access only for a task that needs it, and know which tools can delete. Tella documents a delete tool, Claap documents that it has none, and Clipy's hosted server is read-first by design, with no delete tool.
Which MCP recorder fits bug reports, UI feedback, demos and meetings?
Clipy is the pick for every use case below: browser bug reports, bugs your customers record, UI feedback, desktop app bugs, product demos, and sales and customer calls.
| Use case | Pick | Why | Others in this space |
|---|---|---|---|
| Browser bug reports | Clipy | Console logs, network requests, GraphQL errors and labelled clicks from the Chrome extension, with no 5-minute cap and a link any agent can read | Jam (5 minutes on free) |
| Bugs your customers record | Clipy | Recording request links land in your inbox, and the share link plays with no account for people and agents alike | Zight request links (paid plans) |
| UI feedback ("change this, move that") | Clipy | Key moments pair the frame with what you pointed at, plus click coordinates on Mac app and Chrome extension recordings | Tella, which documents frames and recorded clicks |
| Desktop app bugs | Clipy | The Mac and Windows apps record outside the browser, and agents read the result through the same tools | Cap's CLI records displays and windows |
| Product demos | Clipy | The agent can cut a video by its transcript, and every demo gets chapters, captions, a call-to-action button on Pro and viewer analytics | Tella |
| Sales and customer calls | Clipy | Transcript, summary and Q&A on every recording, plus search inside every transcript across your library | Claap |
What about tools that do not lead with MCP?
Builder Clips takes a different route: share links "designed to be read directly by AI agents", with a JSON map, a timestamped transcript, frames fetched by timestamp, and browser diagnostics captured by its Chrome extension. No MCP server is needed. Do not exclude a useful workflow because its access mechanism has a different name.
For static interface feedback, Agentation offers UI annotation context instead of video. Clipy also works without MCP: a normal Clipy share link plus the Clipy skill lets an agent read a recording's context document directly.
How do you connect an MCP recorder in Claude Code or Cursor?
The pattern is the same for any hosted server: register the URL, then sign in. These are Clipy's documented commands from the MCP docs.
Claude Code
Hosted server with browser sign-in:
claude mcp add --scope user --transport http clipy https://clipy.online/api/mcp
claude mcp login clipy
Hosted server with an API key from your API keys page:
claude mcp add --scope user --transport http clipy https://clipy.online/api/mcp --header "Authorization: Bearer clipy_sk_live_xxx"
Local server:
claude mcp add --scope user clipy -- npx -y @clipy/mcp@latest
Cursor
Add the local server to ~/.cursor/mcp.json:
{ "mcpServers": { "clipy": { "command": "npx", "args": ["-y", "@clipy/mcp@latest"] } } }
Run npx @clipy/cli@latest login once first, so the local server can find your key. Restart Cursor and check that the Clipy tools appear in its MCP settings.
Check that it works
Ask the agent to list your three most recent recordings, then to read the transcript and key moments of one. If it cannot see a recording, check the account you signed in with and the recording's sharing before widening any permission. For other vendors, swap in their URL: Tella's docs, for example, give claude mcp add --transport http --scope user tella https://api.tella.com/mcp.
What should you compare in the output?
Compare what the agent actually receives: the transcript, visual context, pointer or element context and browser diagnostics, plus permissions, tool actions, capture direction, and plan and setup.
| Evidence or behavior | Question to ask |
|---|---|
| Transcript | Does it preserve the relevant narration and timestamps? |
| Visual context | Can the agent inspect the actual state at the important moment? |
| Pointer or element context | Was the target captured, or is the agent inferring it? |
| Browser diagnostics | Are relevant events present for this capture source and mode? |
| Permissions | Can the intended agent retrieve the recording without widening access? |
| Tool actions | Which operations read data, and which can change it? |
| Capture direction | Is a person recording for the agent, or the agent recording verification? |
| Plan and setup | What are the actual usage limits and one-time configuration steps? |
How do you test a useful recording-to-agent handoff?
Record a small bug with a known expected behavior, have the agent retrieve it through the intended integration, and check that it separates what it observes from what it infers. Then verify any change against the original failing interaction.
- Choose a small bug with a known expected behavior and safe sample data.
- Record the starting state, the actions and the result.
- Retrieve the recording through the intended integration and account.
- Ask the agent to separate what it observes from what it infers.
- Correct misunderstandings before asking for a code change.
- Verify the change through the original failing interaction and appropriate tests.
For example, a menu covering a button might be a layout issue, a viewport issue or an expected modal state. The recording locates the symptom. The expected behavior and code inspection still matter.
Use the Claude Code walkthrough for a human-to-agent bug handoff. Use recorded verification when the agent is producing evidence after a change.
What are the common questions?
Does every agent-readable recorder require MCP?
No. Builder Clips uses agent-readable share links, and Clipy share links can be read through its skill without MCP. Choose the path your agent supports.
Is Clipy MCP read-only?
By default, yes on the hosted server. You can grant optional organize access for renaming, tags, folders and sharing. The local server can also record when its key has the Record and upload scope. The hosted server has no delete tool, by design.
Can an MCP server give an agent vision of the video?
Some can, through still frames rather than a video stream. Jam's getFrames returns frames at timestamps or as an overview grid, and Clipy returns frames at key moments. The agent then reads those images with its own vision model. Servers that expose only transcripts give the agent words, not pictures.
Does every recording contain browser logs?
No. Capture source, mode and processing matter. Inspect the actual output and treat missing evidence as a gap.
Which MCP screen recorder is free?
Jam and Claap list MCP on their free plans. Cap's CLI and MCP server are free, though transcripts and AI need Pro. Clipy's MCP server is included during its 7-day trial and on Pro. Zight requires a paid plan, and Tella has no free plan.
Can I connect a screen recording MCP server to ChatGPT?
Some, yes. Clipy's MCP docs describe adding the hosted server as a custom app with OAuth where your ChatGPT plan and workspace allow developer mode, and Tella's docs describe a ChatGPT connection too. Check each vendor's docs for the clients it supports.
Can an MCP server delete my recordings?
Only if the server has a delete tool and you approve it. Tella documents a delete_video tool that moves a video to trash. Claap says its MCP server has no delete tool, and Clipy's hosted server is read-first by design, with no delete tool. Grant write access only when a task needs it.
Is this a benchmark ranking?
No. It is a documentation-based comparison, checked against each vendor's pages on October 9, 2026. It does not report head-to-head performance measurements.
Start with one real task. Open Clipy for agents, connect the server and evaluate the evidence your agent actually gets before changing your team's workflow.