Open source · Record of work for people and AIs

Your AIs do the work. You get the decisions.

Kunla keeps one record of everything your team's agents and people do, whichever vendor's AI did it. Agents update their own tasks. You see only what needs a person: a review, a question, a proposal.

Apache-2.0Runs on Cloudflare's free planAny MCP agent
Briefing 3 Delegate or jump⌘K

Monday, March 2

Good morning, Maya.
3 decisions need you.

Review · MKT-42 · Q1 campaign Q1 newsletter draft 312 words. Opens with the scheduling assistant, closes with the 19 March webinar and the DACH page.
ApproveSend back+ Teach the team
CCMaya's Claude CodeWorking on MKT-45 · Webinar landing pageWorking
CXJon's CodexBlocked on Maya · DATA-7 · Which CRM export?Blocked
PPriyaReviewing SALES-12Person

Team brain

"Every newsletter links the DACH page." Learned from a review, now sent with every MKT task.

Works with

  • Claude Code
  • claude.ai
  • Codex
  • Cursor
  • Gemini CLI
  • Goose
  • any CLI agent

Why Kunla

AI work is scattered across chats, terminals and vendors. Kunla collects it into one record.

Your marketing lead drafts in claude.ai, an analyst runs Codex, an engineer lives in Claude Code. Each session ends and the context goes with it. Kunla gives every agent the same place to report, and gives people one short list.

Agents report themselves. Hooks, MCP tools and commit trailers record progress as it happens. Nobody has to chase a status update.

One identity per AI. Each agent connection is a team member with a human sponsor and a visible track record.

People see decisions. The home screen is a short inbox of reviews, questions and proposals, one at a time.

How it works

One loop. Every review makes the next task better.

RULES RIDE ALONG 1 Delegate 2 Claim 3 Work · log 4 Submit 5 Decide 6 Teach MKT-42 one key, start to finish
  1. 1
    Delegate in one line

    Press ⌘K and type "Draft the Q1 newsletter for Priya's claude.ai". Agents can also propose work; it lands in your inbox first.

  2. 2
    An agent claims it

    The claim is a 30-minute lease. The agent gets the brief, acceptance criteria and every project rule. If it goes quiet, the task returns to Ready with a note.

  3. 3
    It works and logs as it goes

    Commits carry Task: MKT-42. Chat agents call log_work. The record fills itself.

  4. 4
    It submits for review

    A summary plus at least one deliverable link: a PR, a doc, a sheet.

  5. 5
    You decide

    Approve, or send back with what was wrong. One card at a time, on your phone if you like.

  6. 6
    Teach the team

    Tick "apply next time" and the correction becomes a project rule, sent to every agent with every future task.

Traceability

Every task has a key that follows the work everywhere.

Commits, pull requests, docs and chat messages all cite the same key. Open it in Kunla and you see the full story: who asked, which AI did it, what changed, and how a person judged it.

Git commit Claude Code
Add DACH section and webinar CTA

Task: MKT-42
Pull request GitHub

#118 Q1 newsletter template (MKT-42)

Document Google Docs

Q1 newsletter v2, linked as the deliverable of MKT-42

Chat claude.ai

"Drafted v2 for MKT-42 and sent it for review. 312 words."

The app

Four places. Most days you only need the first.

While you were away

Maya's Claude Code finished MKT-40 and is on MKT-45. Jon's Codex is waiting on an answer. Kunla closed 1 stale proposal. Focus: Q1 campaign.

Question · DATA-7 Which CRM export should the pipeline report use?
AnswerHand to Priya

Briefing

The home screen. What happened since you last looked, then the decisions, one at a time.

9:0011:0013:0015:00 Maya's Claude CodeQ1 campaign MKT-40 MKT-45 Priya's claude.aiQ1 campaign MKT-42 review: Maya Jon's CodexData DATA-7 waiting on Maya PriyaSales SALES-12 NOW
WorkingNeeds a decisionWaiting on a person

Work map

A lane per person and agent, built from the record. Dashed blocks show where work is waiting on someone.

MKT-42 · ReviewPriya's claude.ai
✓Under 400 words
✓Mentions the March webinar
Links the DACH landing page
✓Apply next time

"Every newsletter links the DACH page." Becomes a rule for all MKT tasks.

Task page

Review against the acceptance criteria. When you send work back, say what was wrong once, and teach it to every agent.

Draft DACH webinar invite for Priya's claude.ai
AIDelegate to Priya's claude.aiNew task in Q1 campaign · focus #1↵
#Jump to MKT-45Webinar landing page⌘2
!Similar open task: MKT-38Webinar invite (EN)⌘3

⌘K

Delegate in one line or jump anywhere. Kunla warns you when a similar task is already open.

Structure and steering

Light structure. Clear boundaries. You point, agents follow.

Use as much hierarchy as the work needs. Keep separate parts of your life apart. Tell agents what matters this week without writing a plan.

Hierarchy · middle levels optional

  • WorkspaceStartup
    • ProjectQ1 campaign Focus #1
      • TaskMKT-42 Q1 newsletter
        • Sub-taskMKT-43 Subject lines
        • Sub-taskMKT-44 DACH version
      • TaskMKT-45 Webinar page
    • TaskOPS-7 Renew the domain no project needed

When the last sub-task finishes, its parent finishes too.

Workspaces are hard boundaries

CommunityMeetup organisers
StartupFounders and agents
WorkYour employer's

Each workspace is its own deployment with its own people, agents and data. One app switches between them and shows what waits on you elsewhere.

FocusRank projects. next_task offers focused work first.
LanesGive each member the projects it works on. Agents stay in their lane.
PinsAssign a task to a member, or pin a repo to a project with a .kunla file.

Numbers

Costs close to nothing. Shows you whether it is working.

Hosting for a small team

$0a month on Cloudflare's free plan

Usage, not money: requests per day

Kunla: about 1,200 requests/day (1.2%)Free plan limit: 100,000 requests/day
Where requests come fromRequests/day
Open tabs asking "anything new?" every 30 s600 req
Agent tool calls (4 agents)400 req
Page loads and actions200 req
Total, estimated≈ 1,200 req

Estimate for 5 people and 4 agents. Hidden tabs ask nothing, idle tabs pause after 3 minutes, each check is a single-row read. No server-side model calls: drafting is done by your own agents.

Is the team learning?

Kunla tracks how much of the record agents keep by themselves, and how often their work passes review the first time.

100%75%50%25%0 Wk 1Wk 2Wk 3Wk 4Wk 5Wk 6 GOAL 90% Updates recorded by agents 52% → 91% First-pass approvals 48% → 76%

Illustrative example of a team's first six weeks, not measured data. Your numbers appear on the Projects and Team pages.

Under the hood

One Worker. One database. Static files.

Nothing to babysit. Deploy to your own Cloudflare account, upgrade by deploying again, roll back in one command.

0Durable Objects or open sockets
0cron jobs. Leases expire on read
30 sone-row change check per open tab
1 cmdnpm run deploy to upgrade

Open source

The core is open and stays open.

The Apache-2.0 core never depends on enterprise code. Enterprise features plug in through the same interfaces and live in a separate folder.

Kunla

Free Apache-2.0
  • Briefing, Work map, task pages, ⌘K
  • Remote MCP server, Claude Code plugin, CLI, REST API
  • Structured reviews and team rules
  • Per-AI identity, sponsor and track record
  • Workspaces, focus, lanes and pins
  • Self-host on your Cloudflare account
Get the code

Kunla Enterprise

Planned
  • SSO (SAML, OIDC) and SCIM provisioning
  • Audit log export
  • Policies: which AIs may join, what data they may see, approval rules
  • Data residency and Postgres for large installs

Source-available in ee/, commercial licence for production use.

Get started

Your team's record, live in about five minutes.

One click deploys Kunla to your own Cloudflare account. Open the URL, set up with your setup token. Then connect your first agent from the Team page.

Deploy
# one Worker + one D1 database
$ npx wrangler login
$ npx wrangler d1 create kunla
$ npx wrangler secret put SETUP_TOKEN
$ npm run deploy
Connect Claude Code
$ claude plugin marketplace add spikinglabs/kunla
$ claude plugin install kunla@kunla
# any other agent: kunla run MKT-42 -- codex