AI Development Director
Every agent tool can tell you the run finished. aidd is local mission control that makes the run prove it.
Point it at a codebase, feed it a spec, pick the agent and model you already pay for. aidd plans the backlog, launches bounded runs, blocks “done” until tests, lint, and browser gates pass, and keeps the receipts. Your projects, your keys, your machine. No telemetry.
The field
In 2026 you can rent a cloud agent by the task, stack a kanban board over your terminal agents, or generate immaculate specs. None of them will tell you whether the work is actually done. That gap is the whole reason aidd exists.
Hosted agents bill per task—compute credits, premium requests, token burn—and run in their sandbox, not yours. Every retry costs again, and your code lives on someone else’s box.
aidd instead: runs the agent subscription you already have, on your machine, under hard cost, time, and iteration caps.
Kanban boards, worktree apps, and tmux managers show sessions and diffs. This year’s independent roundups note that none ship automated quality gates or plan backlogs from specs—and several are Mac-only or lost their maintainers.
aidd instead: a feature completes only after tests, typecheck, lint, and browser verification pass. The gate decides, not the vibe.
Spec-driven kits produce beautiful requirements, then hand off to an agent and hope. Drift—plausible code that quietly diverges from intent—is the failure mode everyone names in 2026.
aidd instead: carries the spec all the way through: dependency-ordered backlog, acceptance criteria, gate enforcement, and a decision queue for real judgment calls.
| What you need | Cloud agents | Agent boards | Spec kits | aidd |
|---|---|---|---|---|
| Where the work runs | Their cloud | Your machine | Your repo | Your machine, your keys |
| Spec becomes a backlog | One ticket at a time | You write the cards | Yes, then hands off | Dependency-ordered, with acceptance criteria |
| “Done” means | The agent stopped | You eyeballed the diff | The spec was written | Tests, typecheck, lint, and browser gates passed |
| Cost model | Per-task metering | Free, futures vary | Free CLI | Your existing agent plan, capped per run |
| Run receipts | Partial logs | Session history | None | Commits, iterations, tool activity, cost |
| Windows | Browser only | Mostly Mac-first | Varies | Smoke-tested, first-class |
Category snapshot, July 2026. Cloud agents: Devin, OpenAI Codex cloud, Google Jules, Copilot coding agent, Factory. Agent boards: Vibe Kanban, Conductor, Sculptor, Claude Squad. Spec kits: GitHub Spec Kit, AWS Kiro, BMAD. All of them are fine tools for the slice they cover—aidd covers the loop.
The loop
aidd keeps the work visible: it turns intent into a backlog, selects runnable work, launches bounded agent runs, verifies the result, and parks anything that needs a human call.
Add application roots in Settings. Any folder with an
.aidd/ contract becomes a project,
including foreign stacks.
Onboarding turns a plain-language spec into dependency-ordered feature work with verifiable acceptance criteria.
Use Claude Code, Codex, OpenCode, Kilocode, or OpenAI-compatible endpoints including LM Studio and Ollama.
Features complete only after the gate passes: tests, typecheck, lint, and browser-level verification against the app.
Every run keeps commits, iterations, tool activity, cost, and an honest summary. Unclear work lands in the decision queue.
Receipts
Before release, aidd built and repaired real applications through four different creation lanes. The point was not a demo. The point was to catch rough edges under instrumentation, file the failures, fix them, and leave a public trail.
The public changelog records the fixes. The walkthrough slot is ready for the instrumented write-up when the video lands.
| Lane | App | Stack | Result |
|---|---|---|---|
| Fresh scaffold from a spec | Habit tracker | Bun + TS + React | 25/25 features passing, gate clean, roughly 2.5 h |
| Third-party template init | Kanban board | Vite + React | 23/23 features, 10/10 headless acceptance scenarios in the gate |
| Full-stack template init | SMB infrastructure dashboard | Bun + Elysia + React + Drizzle | 26/26 features, 30/30 gate steps green |
| Existing-codebase ingest | Flask tutorial app | Python + Flask |
Profiled the foreign stack, produced a grounded security
audit, then fixed findings with pytest as the gate. Zero
writes outside
.aidd/.
|
Trust model
The default posture is boring in the places that matter: local data, explicit model backends, bounded runs, and no mystery write surfaces.
The runtime, panel, ledger, and project metadata live locally. Network traffic goes only to the model backends you configure.
No phone-home. The write allowlist is documented, including what
lives under
.aidd/ and when source can change.
Cost caps, iteration caps, wall-clock caps, dirty-tree guards, and parked decisions keep automation from pretending certainty.
Get it
Windows 11 + PowerShell 7 is the current smoke-tested path. Git on PATH is recommended. The standalone zip bundles its own runtime.
# 1. Download the zip + SHA256SUMS.txt from the latest release
# 2. Verify before extracting
$zip = (Get-ChildItem aidd-v*-bun-windows-x64-modern.zip).Name
$want = (Select-String .\SHA256SUMS.txt -Pattern ([regex]::Escape($zip))).Line.Split(' ')[0]
$got = (Get-FileHash .\$zip -Algorithm SHA256).Hash
if ($got -ieq $want) { 'OK' } else { throw 'CHECKSUM MISMATCH' }
# 3. Extract to a stable location, optionally add the folder to PATH
Expand-Archive .\$zip -DestinationPath C:\Tools\aidd
# 4. Start the control panel
aidd-web # then open http://127.0.0.1:3210
aidd ships under
FSL-1.1-Apache-2.0: free for any use except building a competing product or service, and
each version automatically becomes Apache-2.0 two years after release.
The .aidd/ contract format is an open spec
anyone may implement. Spernakit, the companion full-stack template, is
MIT.
Need a competing-use, white-label, or enterprise license? Open a discussion and say what you are building.