Home Yield
Yield
A helper that does tasks, asks before saving anything, and stops at the spending cap you set.
Version 0.5.0. Real runs on AI services are recorded; who ran and checked each one is in the first item not finished.
70 / 100 · Beta on our own rubric · 13 not finished in Yield
This button opens your email app with the subject "Waitlist: Yield". There is no sign-up database yet; your email comes to support@shipitworks.com and we keep it only to tell you when Yield is ready. To be removed, send an email with the subject "Remove me". Privacy notice.
Scorecard
70/ 100
Stage: Beta on our own rubric. Google Play: No Android app.
Primary journey: shown with sample data
Scored 2026-11-08 on our own rubric; repository at commit dfaae6c.
Recorded 2026-11-08. This video shows one journey: a real run on Claude, the Needs-you approval clicked by the recording script (as its caption says), the spending cap, and the length-limit warning on a staged reply from a stand-in model (as its caption says). It does not show every page; the full-product video is in production. The scorecard in the video reads 76; the site shows 70/100 on our own rubric (2026-11-08). Re-recorded 2026-11-08 against agentorchestration-autonomous-aigov-platform main dfaae6c; caption at 1:04-1:17 now reads $0.25, $1.00 or $5.00; added the Needs-you length-limit warning (#9), shown on a staged reply from a stand-in model; the card text and date are updated. Encoded CRF 16, -16.3 LUFS. Voice is AI-generated (ElevenLabs). Shows made-up demo data. Read the transcript.
What we promised, and what we have not finished. We promised one public rubric for every product, a list of what is not finished on every page, and a named source for every number. Most of these products are not finished, and each page says where it stops. Read the rubric or every claim and its source.
Works today
Each line links to the scorecard row that backs it.
- Approval bound to exact bytes, hard + soft spend cap, tamper-evident record
- 32 jobs in six kinds, practice helper (mock) only
- Connect a real AI with encrypted key: Anthropic re-run by the reviewer; OpenAI, OpenRouter, HF, Ollama and xAI recorded by the builder
- MCP server (tools by key type, start task, approval refused)
- Results page (cost, hours freed from user-entered baselines, CSV)
- Helpers per AI: state (working/waiting/stalled/idle), job props, 4 monitor scenes (code, bird, space, maze)
- Click-to-choose a helper + 5-second 'on stage' preview
- Done celebration
- .deb installer
- Results by P&L line, with a worked calculation for every money number (Lane 1b)
- Warn on the approval screen when the AI's answer was cut off at its length limit; strip a whole-answer code fence from CSV/JSON/text before the fingerprint
Not finished
From the same scorecard, next to the claims.
- Real runs for five of the six services were recorded by the builder, not independently verified; the sixth (Anthropic) was re-run on 2026-11-03 by the Claude session that reviewed this project, approved by a person in the app, final status done, cost $0.003969.
- The stage walkthrough was filmed with the states held in turn by the demo harness, not during one live run; no playable game
- native/ is a stub; no independent security audit has passed
- xAI us-central-1 is priced about 10 percent above the table; the on-screen workings use today's price, not the price at the time of the step (proof/proof-pack.json:401)
- gitleaks flags a fake key joined from parts in e2e/d1-adversarial.spec.ts:14 (a false positive); the allowlist is not yet extended (proof/proof-pack.json:61)
- Feature 13 (the cut-off warning and fence stripping) is on main (4108613) with a logged run on a stand-in AI; it has not been seen with a real AI answer. Opus and Fable were not tested against the thinking change; Sonnet 5.5 rejects thinking:disabled (HTTP 400), which is why it is sent to Haiku 5.x only. Commits 48be421 and 3547323 are not on main under those hashes (not verified whether squashed into 4108613).
- A task refused by the hard cap shows 'price not known' and is counted as an unpriced step on Results, although no call was made (seen on camera 2026-11-03). The wording is wrong; not fixed.
- The before/after ROI runs (videos/yield/roi/RESULTS.md) stop at the approval step: no manual-vs-agent comparison exists and approval time was not measured. One Sonnet reply in 18 after-fix runs still reached its 1,024-token limit (flagged on the approval).
- The v5 video (published 2026-11-05) shows one journey: a real run on Claude, the Needs-you approval clicked by the recording script (captioned on screen; not evidence of a person approving) and the spending cap. It does not show every page; the full-product video (v6, BACKLOG.md item 2) is not recorded yet.
- The demo video's 1:04 to 1:17 section shows spend choices of $0.25, $1 and $4; the app now offers $0.25, $1.00 and $5.00 (CHANGELOG.md on main, 4108613). Needs a re-record.
- Hosted CI on main: runs for 4108613 and d55a2e1 succeeded (gh run list, 2026-11-08); the run for dfaae6c was queued when checked, so CI on the current HEAD is not verified.
- Work in lane rounds 8 to 12 (native Linux window, money amounts on Needs you, why a task stopped, What it saved, Download this record and more; PRs #75 to #121) is on open branches, not merged to main, and is not counted in this score.
- Some of its pages still show an old contact email; they need to point to support@shipitworks.com.
What changed
Dated lines from the scorecard. The source code is private, so these lines are the public record.
- 2026-11-08: Re-scored 70/100 on our own rubric (was 76 on the old held rule). The 25 feature points are split over 13 features: one proven by a logged test run (the cut-off warning, PR #9, checked with a stand-in AI, not a real AI answer), ten built and tested with sample data, two not built or not tested yet. Tagged v0.4.0 at main dfaae6c.
- 2026-11-08: Hosted CI now runs on the owner's self-hosted runner (PR #123). The README matches the recorded real runs (PR #10, release v0.3.0); notes-only pull requests #24 and #33 changed no code.
- 2026-11-07: The Needs you card now warns when the AI's answer stopped at its length limit, and a CSV, JSON or text answer is saved without the code fence around it. Tested with a stand-in AI, not with a real AI answer. Release v0.3.0.
- 2026-11-07: The README now matches the recorded real AI runs (release v0.3.0). The demo video's 1:04 to 1:17 section shows the older behaviour and has not been re-recorded.
How Yield compares
Each row says only what that project's own page says. We list tools we respect.
| Product | Asks before acting | Spend limits | GitHub stars |
|---|---|---|---|
| Paperclip | Approval stages (README) | Budgets that pause at limits (README) | 109.6k |
| Omnigent | Policies ask before shell or file writes (README) | Can cap spend (README) | 11.1k |
| AnythingLLM | Not mentioned on its README | Not mentioned on its README | 67.8k |
| Yield | Approval bound to exact bytes, hard + soft spend cap, tamper-evident record | Hard and soft spend cap | Not public yet |
The next step for Yield.
Join the waitlist