"Claude Code writes my code beautifully. I don't need a runtime wrapped around it." You're right. And you don't feel the need for one because you are the runtime: you open the worktrees, recover the dead runs, merge what collides, decide whether "done" is really done. PaellaDoc keeps the product contract alive until the work is validated, with evidence. From the AI conversation that produced the line to the Playwright e2e video that proved it works. On your machine.
Claude Code writes your code beautifully. At the keyboard, it doesn't feel like anything is missing. And you're right.
What you don't see is why it works: because you are the runtime.
You split the product into tasks. You open the worktrees. You keep eight sessions alive, remember which branch is safe, catch the build that's green but quietly wrong, recover the run when it dies. The tool writes the code. You run the loop around it, by hand, without noticing.
The bottleneck isn't typing. It's that loop: context, decisions, recovery, evidence. And the better you are, the easier it is to miss.
Three things the tool doesn't take off your head
No coding tool solves these. You carry them, by hand, until you move them into a layer that holds them: the product contract, the gates, the evidence, and the memory of every run.
Claude Code writes the code.
Codex runs the tasks.
PaellaDoc keeps the contract alive.
Point PaellaDoc at it. Reverse-intake reads your code and seeds PRD, epics, stories.
start here →No repo yet. Write the PRD with the orchestrator. Stack template scaffolds the first commit.
start here →PRD, epics, acceptance criteria. The work-item view shows you what every agent decided.
start here →Workspace mode. Describe what you want, iterate, see evidence. The agent ships; the gate verifies.
start here →Git stops at "who committed it". PaellaDoc keeps the chain: spec → agent conversation → code → e2e proof. Open any line months later, refactor without fear, onboard without ceremony. The thread doesn't break.
CI runs the tests the agent wrote. PaellaDoc runs Playwright e2e against your running app — your dev server, the same one you use — with the agent's code applied, and writes screenshots + video + trace to disk. No evidence, no green light. The agent doesn't decide. The gate does.
Everything on your disk, in plain SQLite. No cloud, no vendor cage. Anonymous telemetry, opt-out anytime — chats, code, prompts and secrets never leave your machine. Bring any model — Claude today, Codex tomorrow, a local Llama next week.
A year goes by. You open a story you barely remember. You see exactly who wrote which line, why, what spec asked for it, what test proved it works, and what depends on it. No archaeology. No "who shipped this?" No abandoned PRs from agents nobody can debrief.
# your orchestratorYOU (1 human) ├── orchestrator · paelladoc-ado · local │ ├── project: acme/portfolio-v3 │ │ ├── worktree-01 → claude · story-014 · ✓ │ │ ├── worktree-02 → codex · story-015 · ↻ │ │ └── …16 more │ ├── project: paelladoc/ado-core │ │ ├── worktree-01 → codex · refactor · ✓ │ │ └── …23 more │ └── project: cabin/local-llm-bench │ └── worktree-01 → llama-local · ✓ (offline) └── context_store · sqlite · ~/.paelladoc/ ├── chats.db 2.4 GB ├── decisions.db 180 MB └── artifacts/ 11 GB
Every agent runs in its own git worktree. Its own branch, its own filesystem view, its own conversation. Cleanup is automatic. Merge wars stop being a thing.
You are not the bottleneck. Queue 50 user stories before bed; wake up to 50 worktrees finished, replayed, and ready to review at your speed.
The orchestrator is on your machine. The context is on your machine. Each AI developer calls out to whichever brain you choose — or to a local model you self-host.
The real value of working with AI for months isn't the code it generates. It's everything you build along the way: your decisions, your edge cases, the way your domain actually works.
Most agentic tools quietly turn that into someone else's training data. Or they lock it inside a single frontier model — switch and you lose everything.
PaellaDoc ADO inverts it. Your context lives in plain SQLite, on your disk, in formats you can read. Point any model at it — Claude today, Codex tomorrow, a local Llama from a cabin next week. The brain changes; your knowledge doesn't.
Most agent stacks let the agent declare its own work done. The PR lands. Three weeks later, your test suite catches the lie — if you're lucky. PaellaDoc inverts it: every acceptance criterion runs as a real Playwright e2e against your running app — your dev server, the same one you use. Screenshots, video, trace — saved to disk before status flips to done. No evidence, no green light.
done when evidence exists and passesAir-gapped clients. On-prem mandates. Regulated industries that forbid frontier APIs. Or just a cabin off-grid. PaellaDoc ADO keeps the product methodology intact — PRD, epics, stories, acceptance criteria, golden-gate validation — and routes each story to whatever brain your context allows. Same hierarchy. Same evidence trail. Same shipping bar.
Day-0 frontier MoE, running locally. The orchestrator drives it like any other backend — same PRD, same golden gate, same audit trail.
Frontier-everywhere is a money problem nobody is talking about. PaellaDoc starts by giving you the receipt: every model call logged with its real cost, attributed to the AC that triggered it, queryable from your terminal. You pick the policy per project — lock to local-only, allow frontier, mix per AC. PaellaDoc enforces it and tells you what it cost.
~/.paelladoc/chats.db · $ paelladoc cost --report
PaellaDoc is open-core. The desktop orchestrator, context graph, memory ranking, router and execution engine are the product core.
The extension surface is open: plugin SDK, manifest schema, CLI adapters, MCP packs, validators and examples.
Build adapters for your favorite agents. Keep your context local. Don't wait for me to support your stack.
Wire any agent CLI — Claude Code, Codex, Kimi, Cursor, your own — into the orchestrator. The router calls them. PaellaDoc runs them.
Bundle Model Context Protocol servers as plugins. Vector DBs, custom tools, internal APIs — exposed to every agent, scoped per project.
Custom golden-gate checks. Type purity, license audit, perf budget — your bar, not mine.
Opinionated bootstraps. Next.js + Postgres, Rails + Sidekiq, your house template — not mine.
Reusable skill recipes. PR review, migration scaffolding, test-pyramid generation — packaged, versioned, shareable.
Boundary by design. Plugins cannot access raw chats, KG internals, embeddings, router scoring or secrets — unless PAELLADOC grants an explicit local permission.
Six side-products. One human. The orchestrator runs each as its own workspace, and the context never bleeds between them.
100 stories across a legacy migration. Queue them, assign brains, watch the tree turn green. The PM sees what every agent decided, not just what it shipped.
Owns the model weights. Owns the context. Codes from a cabin. PaellaDoc just routes the work.
Client code never leaves the laptop. Frontier APIs are opt-in per project. Audit trail in plain SQLite.
Same task, three brains, side-by-side diff. Keep the best output. Stop guessing which model is good at what.
“I'm one person. I run several products. I don't want to be a manager of AI developers — I want to be the architect, and let the orchestrator run.
I also don't want my entire mental model — every conversation, every decision — to live inside a vendor's cloud where it disappears the day I switch tools.
PaellaDoc was for the new paradigm — context and spec-driven over the code itself. The architect writes the spec; the system ships the code. 5× productivity — thesis. 277★ in April 2025, when nobody was talking about this yet.
Now I'm back. PaellaDoc ADO — chasing 100× for one human.”
free for personal use · 100% local · no account · macOS Apple Silicon
v0.2.247 · signed & notarized · verify checksum ↑
Heads up: PaellaDoc is alpha (v0.2). It's signed and safe to run, but it's early and you will hit rough edges. When something breaks, tell me on the forum and I fix it fast.
Windows alpha is available only for trusted testers. It is not code-signed yet, Windows SmartScreen may warn about an unknown publisher, and unsigned alpha builds run at your own risk.
My commitment. The current core features for personal individual use will stay free. Pro, Team, cloud, support, sync or commercial use may require a paid license. read the full promise ↑
team or enterprise? — DM @jlcases on X
