- Building
Pit Crew turns a hackathon project into a narrated pitch video: problem, solution, demo, evidence, what’s next and the ask. You paste a live link or a GitHub repo, or upload app footage, and you get a 1–3 minute video with voiceover, captions and diagrams.
The thinking runs on HarnessRouter. Each video is one task in its own sandbox: the agent explores the site (or clones, installs and runs the repo), writes the pitch and returns a video plan. For GitHub-only projects, team code never runs on our server, only in the sandbox.
Features
- The pitch outline judges asked for, with a script review before anything is voiced
- Live-link recording, plus GitHub-only projects run and recorded inside the sandbox
- Footage mode: upload gameplay or an app recording, and the agent narrates what’s actually on screen
- Statistics checked against their source before they appear
- Guardrails for health and mental-health projects
- 19 production jobs run, 5+

builders’ projects made into videos
- Blocked
-
We sent the agent screenshots as input_image (used to OpenAI Responses). They never arrived, but the request still returned “success” (200) after 24 s, and metadata.ignored_fields came back empty ([]). Nothing anywhere said the image was dropped. The agent guessed what was in the clip.
To reproduce: send /v1/responses a text part plus an input_image (a small JPEG). It returns 200 with ignored_fields: [], and the agent says no image arrived.
What we did: sent images as input_file instead, and it described the frame correctly. About 15 minutes to find and fix, plus a re-run. (See Want #1.)

-
“Too busy” looked like “bad key”
In 2 of the 8 jobs where the agent ran, the model provider hit a rate limit (429) partway through. The task failed with “Your tokenrouter key was refused,” and the work was lost. Our key was fine. Both jobs fell back to a weaker plan (lower scoring pitches).
To reproduce (sometimes): run several sandbox tasks back to back.
What we did: added a 30-second wait and one retry (about 10 minutes), then re-ran both jobs. (See Want #2.) -
An empty credit account looked broken
On our first test, every model returned a plain 403. The real reason, $0.00 credit on the TokenRouter account, only showed up when we called TokenRouter directly. That took about 25 minutes to diagnose. Then we put the new funded key in our own config but forgot to update HarnessRouter’s BYOK settings. A visible key status would have caught that right away.
(See Want #3.) -
No list of what’s in the sandbox
We had to write test scripts to learn what the sandbox has. We assumed ffmpeg was installed, but it wasn’t on the hosted sandbox. We found 4 CPUs, 3.4 GB RAM, 15 GB disk, Node, Python, git, a virtual display and internet access. A 1,500 s timeout was accepted, but we don’t know the real maximum. One probe task alone used about 390k input tokens.
(See Want #4.)
-
- Loved
- A whole agent in one API call. We didn't build an agent loop, tool runner or file handling. One request, and a full agent explores the site, runs commands, records and writes the plan. That's how we shipped so much in one week.
- A fresh sandbox for every job. Each video gets its own isolated computer, so we can safely run other people's repos. It ran and recorded a Next.js app in 154 s without touching our server. This is what made GitHub-only projects possible.
- Files in and out just work. Images we sent as
input_filelanded in the agent's workspace, and the recorded video came back intact through the session files API. - Set up the agent once. Instructions, model and limits live on the Harness, and every job calls it by ID. Our requests stayed tiny, and every job ran the same agent.
- A familiar protocol. UHP follows the OpenAI Responses format, so our code looked like code we'd already written. No new SDK to learn.
Didn't have time to try yet: continuing revisions in the same session (so a change request picks up where the agent left off), and using the live step events to show users the agent's progress. Both are next on our list.
- Want
-
Never drop input silently (Blocked #1)
Either accept input_image (treat it like input_file) or reject it with a clear error like “send images as input_file.” At minimum, list any dropped parts in metadata.ignored_fields. That field already exists, but it stayed empty here.
Helps: anyone sending screenshots for demos, QA or UI review, and anyone porting OpenAI Responses-style code, where input_image is standard. -
Say “rate limited,” not “key refused” (Blocked #2)
Return a clear rate-limit error with a “retry after” time, or retry automatically.
Helps: every long task, so work isn’t thrown away and nobody chases a working key. -
Say “out of credit” when that’s the problem (Blocked #3)
Pass the real reason through, show whether the saved key has credit, and add “fund your TokenRouter key” to the setup checklist.
Helps: every new team, in the first hour, when they decide if the platform works. -
Publish what the sandbox includes (Blocked #4)
A simple page per sandbox type: CPU, memory, disk, max runtime, internet access, installed tools.
Helps: every builder deciding what to run in the sandbox vs. on their own server.
If any of these already exist, we missed them. Please point us to them!
-
Loading