Git-native OKRs
Objectives, key results, check-ins, retros — markdown files in your own repo. Versioned, reviewable by PR, greppable by agents. No SaaS, no lock-in: uninstall is rm -rf okrdev/.
Open source·MIT·Your repo, your files
okrdev is a free, open-source OKR operating framework with an AI coach built in. It installs into any business or project, knows your key results, challenges work that serves none, pre-drafts your weekly check-ins, and parks every shiny idea before it eats your week.
Installed from the plugin directory — no terminal, no marketplace command. Or read the manifesto first.
The deal
The standard arrangement is humans supervising the AI. okrdev adds the mirror image: a coach that knows your mission and this cycle's objectives, and asks the one question that matters before anything gets built. You can always override it — and every override is written down. Freedom, with a memory.
Distraction isn't a character flaw — acting on it the moment it arrives is.
Capture any idea in ten seconds with /okrdev:park. Weekly triage
promotes it, archives it, or sanctions it as a time-boxed side-quest. The good
ideas survive triage. The dopamine doesn't.
The coach pre-writes your weekly check-in from git history and merged PRs — confidence table, what moved, what drifted. Humans add the parts only they know. Wins first, no guilt trips, and a skipped week earns a two-minute catch-up, not an inquisition.
The coach can never block you. It has exactly one power: it writes things down — judgment calls, drift, emergencies that weren't. Accountability without authority. It turns out that's plenty.
Objectives, key results, check-ins, retros — markdown files in your own repo. Versioned, reviewable by PR, greppable by agents. No SaaS, no lock-in: uninstall is rm -rf okrdev/.
The coach classifies work against your KRs, flags drift, and pushes back on vanity metrics and sandbagged targets — at planning, not just at the retro.
With AI filling skill gaps, a marketer can ship a migration and an engineer can run a pricing experiment. Every objective gets exactly one human DRI, end to end.
Every target you push gets a metric watching what you might break — monitored, not targeted. The Goodhart defense, built in.
Optional collaboration rails — KR: tags on PRs, a gate that asks which KR, CODEOWNERS, protected branches — so safety is an environment, not a behavior.
For greenfield projects: Vercel + Neon + GitHub with per-PR preview environments and CI gates, so non-technical builders ship safely because unsafe paths are hard.
Adoption
Frameworks die of over-installation — ceremony before value gets the whole thing deleted by week three. okrdev installs one rung at a time and never adds a level uninvited. Existing codebase? Keep your stack: the method is fully decoupled from it.
Idea capture + weekly triage. Nothing else. If you only ever install one rung, install this one.
Mission, cycle OKRs, weekly check-ins, retros. This is okrdev proper.
KR: tags on PRs, the okr-gate nudge, CODEOWNERS, branch protection — for teams.
The full Vercel + Neon AI-first environment. Greenfield projects, or deliberate migrations only.
The ladder goes both ways: if a level stops earning its minutes, remove its files and drop a rung. A framework you actually run at Level 1 beats one you resent at Level 2.
Quickstart
okrdev is listed in the ChatGPT plugin directory, so the shortest path is a button — no terminal, no marketplace command, nothing to copy — and the same button covers the Codex CLI. Claude Code gets there in a single paste.
One paste
Hand it to any Claude Code session — the desktop app counts — and it runs for you. It uses the CLI you already have and installs nothing.
Set up okrdev for me, in this order:1. Find a claude binary. PATH first; if it is not there, the Claude desktop app ships one kept off PATH. Install nothing. 2. With it: plugin marketplace add backedbydata-co/okrdev then: plugin install okrdev@okrdev 3. Tell me to restart the session.
Click the block to select it all, then copy.
# However you installed it — with your project open, # ask for these by name or in plain language: /okrdev:install # walks the adoption ladder, starts at Level 0 /okrdev:plan # ready for Level 1? draft your first cycle's OKRs
okrdev is one plugin for ChatGPT/Codex and Claude Code — same skills, same files, same eight commands. Only the keystroke differs: Codex reaches them with an @ mention, Claude Code as /okrdev:park. Plain language works in both.
If you already use ChatGPT, the directory button above is the whole setup — no terminal at any point. Claude Code is Anthropic's agent; get the desktop app at claude.com/claude-code, and you can stay out of a terminal there too — hand the paste above to a code session and Claude runs it. It also runs as a CLI and in IDEs, and okrdev works in all of them; ./okrdev/install.sh does the whole thing in one for scripts and CI, where there is no agent to ask.
One okrdev/ directory and one marked, removable block in your agent's instructions file — CLAUDE.md on Claude Code, AGENTS.md on Codex, never both. Level 2 optionally adds a few files under .github/. Your build system, deploy pipeline, and framework choices: untouched.
The skills are plain markdown — copy them into your repo and hand any capable agent the docs. The format is the framework.
The method runs fine in a repo that contains nothing but the okrdev/ directory. OKRs, check-ins, and the parking lot don't care what you sell.
Get help
okrdev is complete on its own — the paid paths buy help and accountability, not features. There is no locked tier.
DIY — do it yourself
Free · MIT-licensed
DWY — done with you
Small cohort · led by Alex Chisholm, okrdev's author
For when the method sticks but the discipline doesn't.
DFY — done for you
Delivered by Tech Guys on Demand — a division of Tech Guys Who Get Marketing.
Fixed-scope audit, then monthly care · priced on a call
Not sure which? Install it yourself first — it's free and takes ten minutes. You can add help later; the framework doesn't change underneath you.
Questions
Yes. MIT license, no feature gates, no telemetry, no account. Your OKRs are
markdown files in your own repo — delete the okrdev/ directory and
okrdev never existed. The paid paths above buy human help, not software.
Never. Blocking is how tools get uninstalled — usually mid-crisis, usually forever. Overrides always work, and every one leaves a line in the week's check-in. Accountability without authority.
The shortest path is ChatGPT: okrdev is in the
plugin
directory, so it's one button and you're installed — then @okrdev
in ChatGPT or Codex. Claude Code takes
one paste — ask any Claude Code session to run
claude plugin marketplace add backedbydata-co/okrdev &&
claude plugin install okrdev@okrdev — after which okrdev is in every session
including the desktop app. Neither? The skills are plain markdown: copy them into
your repo and hand any capable agent the docs.
Your files, and nothing else. okrdev ships skills only — no MCP server, no telemetry, no account, no network calls of its own. It reads your repo and writes markdown into it; installing it grants no third party anything, because there is no third party.
No. The method is fully decoupled from the stack: Levels 0–2 don't care whether you deploy to Vercel or a VPS, use Postgres or a spreadsheet, or even write code. The opinionated stack module is for greenfield projects and deliberate migrations only — it's never a prerequisite.
Cascade theater (no trees of aligned sub-OKRs), set-and-forget plans, metric tyranny (every pushed target gets a health metric watching what it might break), and ceremony — a three-line check-in is valid, and a skipped week gets a two-minute catch-up.
okrdev is built in the open by Alex
Chisholm, and it runs on itself — this cycle's OKRs, check-ins, and parking
lot are in the repo for anyone to read. This page lives there too, under
site/.