A tool an AI agent calls to check the shot before I pay for it

Lola Lens is a small MCP server that an AI agent can call. Before any video is paid for, it checks what a planned camera move shows in a 3D room, so the prompt only names what that move shows in the room.

Experiment

This picture shows how the check works, not the session below. It's one exported frame of the wide view in the ID pass, where every object is one flat colour. A camera frame moves over it, and the list counts how much of the frame each object fills, so its numbers aren't the ones in the evidence.

What happened

On 19 November an AI coding agent used Lola Lens to check one shot from my series, The Life of Lola. It compared a push from 3 views and chose one. Lola Lens found 8 objects the prompt names but the camera doesn't show well, and the agent found a bug in Lola Lens. After I dropped those 8 objects the check passed, and after the fix the move sentence named the bed.

  • 8 of 18 objectsThe room sentence named 18 objects, and the left push doesn't show 8 of them well. 7 are never seen, and the desk chair leaves the frame at 3.2 seconds.
  • Failed, then passedThe turn 1 check failed with 8 issues. The checks in turns 2 and 3 passed after I dropped those 8 objects.
  • 1 bug foundThe agent found that the move sentence said toward the window while the camera aims at the bed. It was fixed before turn 3.
  • 13.97% to 24.5%The bed's share of the frame during the left push, the one the agent chose.

This is one session on one shot, which is a small sample. Lola Lens measured a 16:9 frame, but the take was vertical 9:16, so these shares are for a wider frame than the real clip. The agent's own words, in the details below, are shortened, and I haven't checked them.

The session

Every row is from the saved logs, in the order it happened. Open Show the details to see the exact tool calls, the results and the agent's reports.

The recorded session, one row for each thing that happened, from the first turn to the last
TurnWhoWhat happenedResult
Turn 1Dana

My prompt gave the agent one shot line and asked it to plan and check a camera shot. The bed and the pillows had to stay at 0.5% of the frame or more, and the agent had to stop if a decision was needed.

Show the details

The shot line, episode 1, act 1, take 1, shot 2 (4 to 8 s)

“slow push in across her dark navy bedroom (reference): Lola asleep under the dark grey bedspread in navy satin pyjamas and a champagne satin eye mask, an open laptop glowing on the duvet beside her; she fell asleep coding.”

The prompt the agent was given, in short

Plan and check a camera shot in Lola’s bedroom for this shot line before any video is made. The bed and the pillows must stay in frame, at 0.5% of the frame or more at every sample time. Say which things in the shot line the room does not contain. If check_prompt needs a decision, stop and report the questions, because the decision belongs to the person. Do not generate any image or video.

Result The 0.5% line is set
Turn 1Lola Lens

Lola Lens sent the agent the room layout. It has no objects for Lola, the laptop, the eye mask or the bedspread.

Show the details

get_layout at 0:14 in turn 1, with {}

The result, in short

The room is 4.2 by 4.6 m and 2.7 m high. It has 23 objects, 5 room parts and 18 set and prop objects, and all of them are present. There is no object for Lola, the laptop, the eye mask or the bedspread.

Result 23 objects
Turn 1Lola Lens

For the agent, Lola Lens measured a 5 second push from the wide, left and right views, and drew text maps of the first and last frames of the left push. The right view never shows the pillows.

Show the details

what_it_sees at 0:17 in turn 1, with {"view":"wide","move":"push","samples":9}

The agent's words before the call, not checked by Dana

No Lola, laptop, eye mask or bedspread objects in the layout. Let me compare the three views with a push.

The result, in short

Wide view, push, 5 s, 9 samples. The bed covers 18.56% to 19% of the frame and the pillows 2.26% to 2.95%. The globe lamp and the floor plant are never seen. The nightstand leaves the frame at 1.88 s, and the rug and the dresser at 3.13 s.

what_it_sees at 0:17 in turn 1, with {"view":"left","move":"push","samples":9}

The result, in short

Left view, push, 5 s, 9 samples. The bed covers 13.97% to 24.5% of the frame and the pillows 1.08% to 1.81%. The desk, desk lamp, pencil cup, bookshelf, dresser, corkboard and hanging plant are never seen. The desk chair leaves the frame at 3.75 s.

what_it_sees at 0:18 in turn 1, with {"view":"right","move":"push","samples":9}

The result, in short

Right view, push, 5 s, 9 samples. The bed covers 13.68% to 16.78% of the frame. The pillows are never seen, and neither are the nightstand, globe lamp, wall shelves, photos and floor plant.

peek at 0:41 in turn 1, with {"t":0,"view":"left","move":"push","cols":64,"rows":20}

The result, in short

A text map of the first frame of the left push, 64 by 20 characters, with one letter for each of 15 things in view.

  • # Back wall
  • A Window
  • B Curtains
  • C Bed
  • D Pillows
  • E Nightstand
  • F Globe lamp
  • G Wall shelves
  • H Rug
  • J Desk chair
  • P Photos and string lights
  • Q Floor plant
  • [ Left wall
  • ^ Ceiling
  • _ Floor

peek at 0:42 in turn 1, with {"t":1,"view":"left","move":"push","cols":64,"rows":20}

The result, in short

A text map of the last frame of the left push. 13 things are in view. The desk chair and the ceiling have gone.

  • # Back wall
  • A Window
  • B Curtains
  • C Bed
  • D Pillows
  • E Nightstand
  • F Globe lamp
  • G Wall shelves
  • H Rug
  • P Photos and string lights
  • Q Floor plant
  • [ Left wall
  • _ Floor
Result Left bed 13.97% to 24.5%
Turn 1AI agent

The agent chose the left push and set it to 4 seconds, because the shot line covers 4 to 8 seconds.

Show the details

plan_shot at 0:45 in turn 1, with {"view":"left","move":"push","duration":4,"samples":9}

The agent's words before the call, not checked by Dana

The left push ends with the bed filling the middle of the frame and the pillows clear. I'll set that as the shot. It runs 4 s, because the shot line covers 4–8 s.

The result, in short

Left view, push, 4 s. The camera moves from (1.55, 1.45, 2.5) to (0.923, 1.307, 1.862) and aims at (-1.3, 0.8, -0.4) the whole time.

Result Left view, 4 seconds
Turn 1Lola Lens

The check didn't pass. 7 objects the prompt names are never seen, and the desk chair leaves the frame at 3.2 seconds. Lola Lens asked a question about each one, and in headless mode all 8 came back cancelled.

Show the details

check_prompt at 0:47 in turn 1, with {"view":"left","move":"push","minPct":0.5}

The result, in short

Not passed. 8 issues. 7 objects named in the prompt are never seen, and the desk chair leaves the frame at 3.2 s. The server asked a question about each one, and in this headless run all 8 came back cancelled.

The prompt sentences check_prompt returned

The room contains: window, curtains, bed, pillows, nightstand, globe lamp, wall shelves, rug, desk, desk chair, desk lamp, pencil cup, bookshelf, dresser, corkboard, photos and string lights, floor plant, hanging plant. The camera pushes in slowly and smoothly toward the window. No cuts.

Result Failed8 issues
Turn 1AI agent

The agent stopped and left the 8 decisions to me. It also reported a bug in the move sentence, which is in the row between turns 2 and 3.

Show the details

The agent's report for turn 1, in its own words, shortened, not checked by Dana

I stopped at check_prompt because it needs decisions that belong to you. It sent you a question for each issue, and every one came back cancelled, so all 8 are still open and I have not chosen any answers.

Plan so far: left view, push, 4 s (the 4–8 s slot). This is the only setup where the push really closes in on the bed.

Bed and pillows stay in frame for the whole left push. At the last frame the bed fills the middle of the frame and the pillows are clear.

Open decisions. check_prompt builds the layout sentence from every object in the room. These are named in it but the camera doesn't see them properly:

  • Never seen: desk, desk lamp, pencil cup, bookshelf, dresser, corkboard, hanging plant
  • Leaves the frame at 3.2 s: desk chair

The options for each are drop_from_prompt, mark_not_present, keep or change_move (to still, pan or orbit). Changing the move would mean checking bed and pillows again. You can answer per object, for example "drop all eight", and I'll pass the answers back in and run the check again.

Another issue to fix: the move sentence says "toward the window", but the camera actually aims at the bed (target −1.3, 0.8, −0.4, inside the bed's box). The video model would be told the wrong thing to push toward. The sentence should probably say "toward the bed".

Things the shot line mentions that the room layout does not contain:

  • Lola herself
  • Navy satin pyjamas and the champagne satin eye mask
  • The open, glowing laptop
  • The dark grey bedspread / duvet. It's at most part of the bed object; there's no separate object for it.
  • The "dark navy" colour of the room. The layout has no colour or material information, so this can't be checked.

The video model will get all of these only from your shot line and the reference image, not from the layout sentence. No images or video were generated.

Result 8 decisions left to Dana
Turn 2Dana

I answered “Drop all eight”. The next prompt passed that on, and said this was a new server process, so the agent had to set the shot again.

Show the details

My words

“Drop all eight”

The prompt the agent was given, in short

The prompt passed on Dana’s answer as drop_from_prompt for the desk, desk lamp, pencil cup, bookshelf, dresser, corkboard, hanging plant and desk chair. It said this is a new server process, so set the same shot again first, run the check with Dana’s decisions, and report anything still open, including the move sentence that says “toward the window” while the camera aims at the bed.

The objects I chose to drop

desk, desk lamp, pencil cup, bookshelf, dresser, corkboard, hanging plant, desk chair.

Result 8 objects dropped
Turn 2Lola Lens

The check passed with no issues, after the agent set the same shot again and sent my decisions. Measured again, everything still named stays above 0.5% of the frame, and the floor plant comes closest.

Show the details

plan_shot at 0:09 in turn 2, with {"view":"left","move":"push","duration":4,"samples":9}

The result, in short

Left view, push, 4 s. The same camera path as in turn 1.

check_prompt at 0:13 in turn 2, with {"view":"left","move":"push","minPct":0.5,"decisions":[{"id":"desk","choice":"drop_from_prompt"},{"id":"desk-lamp","choice":"drop_from_prompt"},{"id":"pencil-cup","choice":"drop_from_prompt"},{"id":"bookshelf","choice":"drop_from_prompt"},{"id":"dresser","choice":"drop_from_prompt"},{"id":"corkboard","choice":"drop_from_prompt"},{"id":"hanging-plant","choice":"drop_from_prompt"},{"id":"desk-chair","choice":"drop_from_prompt"}]}

The result, in short

Passed, with no issues. The 8 objects are left out of the layout sentence.

The prompt sentences check_prompt returned

The room contains: window, curtains, bed, pillows, nightstand, globe lamp, wall shelves, rug, photos and string lights, floor plant. The camera pushes in slowly and smoothly toward the window. No cuts.

what_it_sees at 0:21 in turn 2, with {"samples":9}

The agent's words before the call, not checked by Dana

check_prompt now passes with no issues. I'll check the frame shares once more on this server process.

The result, in short

Left view, push, 4 s, 9 samples. The bed covers 13.97% to 24.5% of the frame and the pillows 1.08% to 1.81%. The lowest share of anything still named is the floor plant, 0.54% at 4 s.

The agent's report for turn 2, in its own words, shortened, not checked by Dana

With Dana's "drop all eight" applied, the plan check passes: check_prompt returns ok: true with no issues. One problem remains: the move sentence says the camera pushes toward the window, but it actually aims at the bed.

Move: push, 4 s (shot 2's 4–8 s slot). The camera goes from (1.55, 1.45, 2.50) to (0.92, 1.31, 1.86) and aims at (−1.3, 0.8, −0.4) the whole time. That point is inside the bed.

Bed and pillows stay in frame the whole time. I checked again on the new server and the numbers matched last turn:

  • Bed: 13.97% → 24.5% of the frame.
  • Pillows: 1.08% → 1.81% of the frame.

Everything still named in the layout sentence stays above 0.5% at every sample. The closest is the floor plant, at 0.54% in the last frame. It still passes, but a slightly longer or tighter push could take it below 0.5%.

The move sentence says the wrong target. It says "toward the window", but the camera aims at the bed. The window is in frame (about 10%), but it is upper right and isn't what the push closes in on. Your shot line says "push in across her bedroom" toward the sleeping Lola, so the generated sentence contradicts both the camera path and your intent. None of the lola-lens tools can change this sentence. Either change it by hand to something like "…toward the bed", or fix how the Lola Studio test script picks the target name. This is Dana's call, so I haven't changed anything.

Things in the shot line that the room layout doesn't contain:

  • Lola herself
  • Navy satin pyjamas
  • The champagne satin eye mask
  • The open, glowing laptop
  • The dark grey bedspread / duvet, which at most is part of the bed object
  • The dark navy room colour, which the layout can't check because it has no colour information

The video model will get these only from your shot line and the reference image.

No image or video was generated.

Result PassedFloor plant 0.54%
Turn 2 to 3Not in the logs

In turn 1 the agent found a bug. The move sentence said toward the window while the camera aims at the bed. In turn 2 it reported that no Lola Lens tool could change that sentence. Lola Lens was fixed between turns 2 and 3. The push sentence now names the object the camera aims at, instead of always saying “toward the window”.

Show the details

The move sentence before the fix, in turns 1 and 2

The camera pushes in slowly and smoothly toward the window. No cuts.

The move sentence after the fix, in turn 3

The camera pushes in slowly and smoothly toward the bed. No cuts.

Result 1 bug fixed
Turn 3Lola Lens

The check passed with no issues, and the move sentence now says toward the bed. My prompt for this turn said the server had changed and that my decision to drop all eight still stood. The agent reported that nothing in the check was still open.

Show the details

The prompt the agent was given, in short

The prompt said the server had been changed so the push sentence names the object the camera aims at, and that Dana’s decision to drop all eight still stands. It asked the agent to set the same shot again, run the check with Dana’s decisions and report whether anything is still open.

plan_shot at 0:09 in turn 3, with {"view":"left","move":"push","duration":4,"samples":9}

The result, in short

Left view, push, 4 s. The same camera path as in turns 1 and 2.

check_prompt at 0:12 in turn 3, with {"view":"left","move":"push","minPct":0.5,"decisions":[{"id":"desk","choice":"drop_from_prompt"},{"id":"desk-lamp","choice":"drop_from_prompt"},{"id":"pencil-cup","choice":"drop_from_prompt"},{"id":"bookshelf","choice":"drop_from_prompt"},{"id":"dresser","choice":"drop_from_prompt"},{"id":"corkboard","choice":"drop_from_prompt"},{"id":"hanging-plant","choice":"drop_from_prompt"},{"id":"desk-chair","choice":"drop_from_prompt"}]}

The result, in short

Passed, with no issues. The move sentence now says toward the bed.

The prompt sentences check_prompt returned

The room contains: window, curtains, bed, pillows, nightstand, globe lamp, wall shelves, rug, photos and string lights, floor plant. The camera pushes in slowly and smoothly toward the bed. No cuts.

The agent's report for turn 3, in its own words, shortened, not checked by Dana

The plan check passes and nothing in it is still open. With Dana's "drop all eight" applied, check_prompt returns ok: true with no issues, and the move sentence now says "toward the bed".

Move: push, 4 s (shot 2's 4–8 s slot). The camera path is the same as last turn: it goes from (1.55, 1.45, 2.50) to (0.92, 1.31, 1.86) and aims at (−1.3, 0.8, −0.4), which is inside the bed.

Frame shares: I didn't run what_it_sees again this turn. The path is identical to last turn's, when the bed went from 14.0% to 24.5% of the frame and the pillows from 1.08% to 1.81%. Everything named in the layout sentence was above 0.5% at every sample. The closest was the floor plant, at 0.54% in the last frame.

Still not in the room layout: the shot line mentions these, but the layout has no objects for them, so the video model will get them only from your shot line and the reference image:

  • Lola herself
  • Navy satin pyjamas
  • The champagne satin eye mask
  • The open, glowing laptop
  • The dark grey bedspread / duvet, which is at most part of the bed object
  • The dark navy room colour, which the layout can't check because it has no colour information

No image or video was generated.

Result PassedToward the bed

The evidence

The three views the agent compared, the last frame of the push it chose, the prompt sentences before and after, and a check of the ray cast against real renders.

Each view with a 5 second push and 9 samples, from turn 1. The share is how much of the frame each object covers. The chosen push was later set to 4 seconds, and in turn 2 the bed and pillows gave the same ranges.
ViewBedPillowsResult
WideBed 18.56% to 19%Pillows 2.26% to 2.95%The bed and the pillows stay in frame, but the bed's share hardly changes. The nightstand leaves the frame at 1.88 seconds, and the rug and the dresser at 3.13 seconds.
LeftBed 13.97% to 24.5%Pillows 1.08% to 1.81%Chosen by the agent. The bed and the pillows stay in frame, and the bed's share grows the most during the push.
RightBed 13.68% to 16.78%Pillows Never seenThe pillows are never in frame, so this view misses what the prompt asked for.

The last frame of the left push, from the peek tool's text map of 64 by 20 characters, drawn here with one cell for each character in each object's ID colour. The bed is C and the pillows are D. 13 things are in view, and the desk chair and the ceiling have gone.

  • # Back wall
  • A Window
  • B Curtains
  • C Bed
  • D Pillows
  • E Nightstand
  • F Globe lamp
  • G Wall shelves
  • H Rug
  • P Photos and string lights
  • Q Floor plant
  • [ Left wall
  • _ Floor
Show the text map

Turn 1, before my decisions and the fix

The room contains: window, curtains, bed, pillows, nightstand, globe lamp, wall shelves, rug, desk, desk chair, desk lamp, pencil cup, bookshelf, dresser, corkboard, photos and string lights, floor plant, hanging plant.

The camera pushes in slowly and smoothly toward the window. No cuts.

Turn 3, the prompt sentences that passed

The room contains: window, curtains, bed, pillows, nightstand, globe lamp, wall shelves, rug, photos and string lights, floor plant.

The camera pushes in slowly and smoothly toward the bed. No cuts.

The ray cast against two rendered frames

Checked against 2 exported ID-pass frames, every object's share is within half a percentage point. The largest differences are 0.45 percentage points for the curtains in the first frame and 0.38 for the back wall in the last frame. Small objects differ more for their size. The desk lamp is 0.19% in the ray cast and 0.34% in the first frame. These two frames are from the wide push, not the left push the agent chose, and they're the only frames I could check.

The dollhouse in Lola Studio exported the wide push in the ID pass, where every object is one flat colour. Press the button to measure each exported frame in this page and compare it with the ray cast.

The first frame of the wide push in the ID pass, rendered by the dollhouse in Lola Studio, with one flat colour for each object
The first frame of the exported wide push, ID pass.
The last frame of the wide push in the ID pass, rendered by the dollhouse in Lola Studio, with one flat colour for each object
The last frame of the exported wide push, ID pass.

Try the tool yourself

This is the real Lola Lens server and a real MCP client, both running in this page. There's no AI and no backend in it, so you press the buttons an agent would use to call the tools. Press Start the server, then press Check the prompt. The server then asks you what to do about each object the camera doesn't show well.

Pressing the button downloads the Lola Lens server, an MCP client and the 3D room of Lola's bedroom. Nothing is downloaded for the demo before then.

How the session was run

The session used the same server as the demo, run as a local process. There's no public repository or install command for Lola Lens, but the demo above runs the same server in this page. These are the steps I followed.

  1. I started Lola Lens as a local MCP server over stdio, and gave a command-line AI coding agent that server as its only MCP server.
  2. I ran the agent headless, with one prompt for each turn. Each turn started a new server process, so nothing carried over between turns.
  3. I gave it the shot line and a prompt with the same requests as the shortened turn 1 prompt below. Because the server was new each turn, every later prompt asked the agent to set the same shot again with plan_shot before it ran check_prompt.
  4. I read the server's questions in the agent's report, and passed my answers in the next turn's prompt, so the agent sent them to check_prompt as decisions. The Limits section explains why the questions couldn't be answered in the form.

The turn 1 prompt, in short

Plan and check a camera shot in Lola’s bedroom for this shot line before any video is made. The bed and the pillows must stay in frame, at 0.5% of the frame or more at every sample time. Say which things in the shot line the room does not contain. If check_prompt needs a decision, stop and report the questions, because the decision belongs to the person. Do not generate any image or video.

The check_prompt call in turns 2 and 3

check_prompt {
  "view": "left",
  "move": "push",
  "minPct": 0.5,
  "decisions": [
    {"id":"desk","choice":"drop_from_prompt"},
    {"id":"desk-lamp","choice":"drop_from_prompt"},
    {"id":"pencil-cup","choice":"drop_from_prompt"},
    {"id":"bookshelf","choice":"drop_from_prompt"},
    {"id":"dresser","choice":"drop_from_prompt"},
    {"id":"corkboard","choice":"drop_from_prompt"},
    {"id":"hanging-plant","choice":"drop_from_prompt"},
    {"id":"desk-chair","choice":"drop_from_prompt"}
  ]
}

The last tool call came 47, 21 and 12 seconds into turns 1, 2 and 3. The prompts for all three turns are shortened on this page. The full prompts are kept with the raw logs.

Limits

  • The frame Lola Lens measures is 16:9, 1280 by 720, like the dollhouse export. The take was vertical 9:16, so the shares are for a wider frame than the real clip, and some objects may be out of the real frame.
  • It checks the plan, not a generated video, and it can't say whether a video model will draw each object correctly. The check step in the Lola Studio test script is planned, not built yet.
  • The ray cast was checked against 2 rendered frames only, the first and last frames of the wide push. No frame of the left push was checked.
  • It was tested on one session, with one shot from one take of episode 1, in one room, Lola's bedroom.
  • The room has no objects for Lola, the laptop, the eye mask or the bedspread, so Lola Lens can't check those parts of the shot line. The agent also noted that the layout can't check the dark navy colour of the room.
  • Because the room has no object for Lola, the shares don't count what she would cover in the video. In the shot line she is asleep in the bed, so the bed's real share would be smaller than the share Lola Lens measured.
  • In headless mode the elicitation questions came back cancelled, so my decisions were passed in the next turn's prompt, not through the question form.
  • Visibility comes from a grid of 96 by 54 rays, so one ray is about 0.02% of the frame and every percentage is approximate. The floor plant's lowest share was 0.54%, in the last frame of the left push. That's 28 rays, and the 0.5% line is 26 rays.
  • The 0.5% line comes from my turn 1 prompt, which asked for the bed and the pillows to reach at least 0.5% of the frame at every sample time. The agent passed 0.5% to every check in this session as the smallest share that counts as seen, and no other value was tried.

How the server works

Lola Lens loads the 3D layout of Lola's bedroom and moves a camera along the planned path. At each sample time it fires one ray through each cell of a grid and records which object each ray hits first. It compares those shares with the prompt, and asks the person before anything changes.

It builds the layout sentence the same way the Lola Studio test script does, so it checks the exact sentence that script would send. It has seven tools and four resources. Three resources can be subscribed to, the layout, the current shot and the prompt, and the server says when one changes.

The ray cast was compared with two rendered frames, which are in the evidence above.

The tools

ToolTakesReturns
get_layoutNothing.The room size, and every object with its id, kind, ID colour and bounding box in metres.
edit_objectAn id, and remove, restore, move or resize with a value.The object, what it now overlaps, and any warnings. A change that leaves the room is refused.
plan_shotA view (wide, left or right), a move (still, push, pan or orbit), and optionally a duration.The camera path, from the same camera maths the Lola Studio dollhouse uses.
what_it_seesOptionally a view, a move, a number of samples and a ray grid size.Each object's share of the frame at each sample time, and when it's first seen, last seen and leaves.
peekA time from 0 to 1, and optionally a view, a move and a grid size.A text map of the frame with one letter for each object.
check_promptOptionally a view, a move, the smallest share that counts as seen, and decisions.The prompt sentences, the issues, and a question for each issue through MCP elicitation.
save_worldA label.A new timestamped file of the edited room. It never overwrites the original.

Next question

Could Lola Studio refuse to spend anything on a shot until the check finds no issues left?

More experiments

Lola's house

Building Lola's house as real 3D rooms

Lola is the lead in my AI video series, and her house only existed as a few pictures that didn't agree. This is the whole attempt to turn them into rooms a camera can move through. There's a kitchen grown from one picture, a floor plan that took 48 versions, and a bedroom built in 3D. It covers what went wrong and what fixed it.

Stereo video

Turning a flat AI film into 3D

I turned scenes from The Life of Lola into 3D by estimating depth and redrawing each eye. Then I measured the same method on a separate test shot, where true stereo could be rendered from a 3D layout of Lola's bedroom.

Lola Studio

Catching mistakes before they're filmed

The Life of Lola is my own series, made with AI video, and every bad take had to be paid for and made again. I started building Lola Studio to help me direct episode 1, and I logged every take along the way.