Open KT is in private pilot→ Join the pilot
Open KT
Join the pilot →

The Shared Intelligence Layer

Your team's AI shouldn't
start from zero.

One shared brain for every coding agent on your team. They inherit yesterday's decisions, share today's discoveries, and skip the 47k-token repo scan every morning.

  • ▸One brain across harnesses
  • ▸Day-one ready agents
  • ▸~99% less re-exploration
Join the pilotWatch a sample run →

Every team will swap Cursor for Claude Code for Codex three times this year. What stays — what compounds — is what the team has already learned. That layer is the product.

The cost of starting from zero

Your agents pay rent every morning,
on knowledge they already had.

  • ▸A new chat opens. The agent re-scans 40 files to find a fix your teammate's agent shipped yesterday.
  • ▸The keep-alive setting, the auth regex, the deploy incantation — none of it survives the chat window.
  • ▸Switch from Claude Code to Codex next quarter — the rest goes with it.

The harness is rented. The intelligence should be owned.

What every session costs you
  • ~47ktokensspent re-exploring per session
  • 3×/ yearharness migrations wipe your context
  • 80%lostof learnings die in chat windows
  • 0%by defaultof it compounds for the team

People who saw this coming

At a glance

Same prompt. Different bill.

WithoutWith Open KT
  • Per session×47k tokens re-explored✓300-token recall
  • First useful action×~6 minutes✓~5 seconds
  • Token bill / month×Scales linearly with sessions✓Sub-linear — recall caps growth
  • Onboarding a new agent×Hours of context-loading✓Inherits the team brief on turn 1
  • When you switch harnesses×Context dies with it✓Memory + skills travel
  • Team capital×Dies in chat windows✓Compounds on disk
  • Audit trail×None — try git blame✓Receipt per recall

Solution · Three layers

Three layers of compounding intelligence.

Throwaway chat becomes durable team capital — searchable, scope-aware, traceable to the session that earned it.

[ 01 ]Memory

Discoveries that compound.

  • ▸One agent finds the answer; every agent after recalls it.
  • ▸300-token lookup, not a 47k-token rescan.
  • ▸Scoped: personal → project → org. Promotion is reviewed.
On disk · examples
  • STAGING_FIX_01

    Postgres drops staging connections without a 60s keep-alive.

    ▸ 14 replays this week · 328k tokens saved · 0 reopens

  • AUTH_REGEX_NOTE

    JWT subject claim is base64url, not base64. Strip padding.

    ▸ 9 replays · ended a 3-day cross-team bug loop

  • DEPLOY_INCANT

    Cloud Run revision must pin --concurrency=8 for our queue worker.

    ▸ 6 replays · 0 outages in the 8 weeks since promoted

[ 02 ]Skills

Procedures every agent inherits.

  • ▸Authored once by your lead — every harness on the team ships day-one ready.
  • ▸L1 metadata loads at startup; L2 expertise pulls only on invoke.
  • ▸Versioned, scoped, replaceable — never a copy-paste prompt.
On disk · examples
  • TAILWIND_STYLE_COP

    Audits component naming, contrast, focus, prefers-reduced-motion.

    ▸ Run by every UI agent · 0 contrast regressions in 6 weeks

  • PR_DESCRIBE

    Writes PR title + summary + test plan in your team's house style.

    ▸ 44 PRs this month · review-ready on first push

  • MIGRATION_REVIEW

    Flags non-backwards-compatible schema diffs before they ship.

    ▸ Caught 3 silent breakages last quarter · zero post-merge rollbacks

[ 03 ]Workflows

Routine work at code speed.

  • ▸Stable patterns get promoted to deterministic ADK pipelines.
  • ▸The LLM shapes intent — the pipeline runs the mechanical part.
  • ▸Stop paying LLM rates for sequences that never vary.
On disk · examples
  • AUTO_RELEASE_SWEEP

    Bumps versions, regenerates docs, opens release PR, posts changelog.

    ▸ 60-min ritual → 90s deterministic run · runs every Friday

  • ADK_DEPLOY

    Builds, runs migrations gated by health checks, promotes if green.

    ▸ 8 deploys/week · 0 prod rollbacks since adopted

  • INCIDENT_TRIAGE

    Pulls logs, paginates Sentry, drafts the incident doc skeleton.

    ▸ First-pass postmortem in 90s · saves the on-call 30 min

How it works

Observe → Promote → Inherit.

One agent's discovery becomes an org-wide capability. Every transition leaves a receipt. Every receipt seeds the next observation.

01

Observe.

  • ▸Watches every session — recalls, skill runs, closed tasks.
  • ▸Mines the patterns your team keeps repeating.
02

Promote.

  • ▸Useful personal memory → project memory.
  • ▸Repeated prompt sequence → candidate skill.
  • ▸Stable skill → deterministic workflow.
  • ▸Reviewed before it escalates. Every promotion leaves a receipt.
03

Inherit.

  • ▸Next session opens with the relevant org context attached.
  • ▸Picks up where a teammate's agent left off.
  • ▸Routine sequences offload to the pipeline. Each loop costs less.

Sample run · the gift

One discovery. Free for every agent that follows.

Same prompt, two agents. The naïve one re-explores from zero; the Open KT agent inherits the team's last session. Watch the cost.

naive-agent · no Open KTWithout Open KT
userPrepare staging for the new deploy.
agentReading 40 files to find staging config…
↳ tokens consumed: 47,184
agentTrying connection — drops after 38s.
agentRe-checking middleware…
↳ tokens consumed: 71,902
warnRe-deriving keep-alive setting from scratch.
agentDone. ~6 minutes.
↳ session cost: 71,902 tokens · 6m 14s
↳ outcome: re-discovered yesterday's fix
agent · openkt-injectedWith Open KT
userPrepare staging for the new deploy.
opnkt[ injecting STAGING_FIX_01 · scope: ORG ]
agentApplying keep-alive (60s) from prior session.
agentValidating connection — stable.
↳ tokens consumed: 412
ok Pre-briefed. 0 redundant exploration.
agentDone. ~5 seconds.
↳ session cost: 412 tokens · 5s
↳ outcome: inherited STAGING_FIX_01 · 99.4% saved
−99.4%cost
6m → 5swall clock
1receipt on disk

Same prompt, two agents. The naïve one rebuilds context from scratch. The Open KT one inherits yesterday's answer and gets to work.

One install · every harness · one shared brain

Claude CodeClaude DesktopCodexCursorOpenCodeAiderHermesAuggieOpenClawAny MCPClaude CodeClaude DesktopCodexCursorOpenCodeAiderHermesAuggieOpenClawAny MCPClaude CodeClaude DesktopCodexCursorOpenCodeAiderHermesAuggieOpenClawAny MCP

Install · MCP over HTTP

One line per harness. Same shared brain.

Claude Code

CLI
$claude mcp add --transport http openkt https://api.openkt.ai/v1/mcp -s user

One command. Verifies connection, stores at user scope so every project picks it up.

Pilot · Limited spots

Starting private pilots.
Be one of the first ten.

We're picking 5–10 design partners — teams who feel the cost of starting from zero and ship fast. Same-week response.

What's included
  • 01
    Direct line to the founders
    Slack or Telegram, same week onboarding. No support tickets.
  • 02
    We ship your first skill with you
    Pair on a custom skill or workflow that fits your team's stack.
  • 03
    Your data, your region
    Self-hosted, region-pinned, or on our infra — your call.
  • 04
    Lifetime price-lock
    What you sign up for is what you pay. Forever, no seat creep.
pilot application · open-kt-pilot.form

5–10 spots · same-week response · no spam

Frequently asked

The objections every CTO has, answered.

Six things people ask before they install. If something else is on your mind, write us.

  • 01Where does the data live? Is anything sent to a third party?

    Your data lives where you choose — self-hosted, region-pinned, or on our infra. Nothing is sent to a third party unless you wire it. You own the store; we own the pipe.

  • 02How is this different from putting context in CLAUDE.md / cursor rules?

    Static files are write-only. They don't learn from sessions, don't scope to projects, and don't compound across agents. Open KT is a living system — it observes, promotes, and inherits.

  • 03What if I switch harnesses next quarter?

    Your intelligence layer stays. Switch from Claude Code to Codex to Cursor — the memory, skills, and workflows travel with you. The harness changes; the brain doesn't.

  • 04Who can see what? How does scoping work?

    Three scopes: personal, project, org. Memory written in one scope only surfaces in that scope. Promotion across scopes is reviewed, not automatic.

  • 05Is there an audit trail for what the agent did?

    Yes. Every recall and every promotion writes a receipt — which agent, which session, which scope. Replayable, exportable, query-able.

  • 06What happens to memories that are wrong or outdated?

    Memories decay. Low-use patterns fade. Wrong patterns get flagged and retired. The system cleans itself; you don't have to.

Ready to compound?

Your team already has context.
Your agents should too.

5–10 design partners. Same-week response. Lifetime price-lock.

Join the pilotEmail us →