BetaNow in private betaThe harness of harnesses, now in private beta

Every agentic harness, working as one team.

OrkiLab is one workspace for Claude Code, Codex, Cursor and every agent you use. They share context, hand work to each other mid-task, and every line they write is traced back to the agent and person behind it.

Fix double loyalty discountcheckout-service

Loyalty discounts apply twice on bundled items. Fix it, add a regression test, and update the pricing docs.

Worked for 12s

Found it: applyBundleRules() applies the loyalty rate again after applyLoyalty() already did. I'll fix it here and hand off the test and the docs so they run in parallel.

Delegates2 working1 working · 1 ready to merge1 merged · 1 working

Regression test for bundlesorkilab/chat/4f2a9c1e+42 −0WorkingMergeMerged

Update the pricing docsorkilab/chat/b81d07e3Working

Codex's test fails on main and passes with the fix. Merging its patch now; I'll open the PR as soon as Cursor's docs land.

Message Claude Code… (/ for commands and skills, ! for shell)

Default model

Works with the agents your team already uses

12 agents supported today: 5 built in, 7 ready-made ACP presets, and any other agent that speaks ACP.

Claude Code
Codex
Cursor
OpenCode
Google Antigravity
Devin CLIACP
Grok BuildACP
DeepSeek HarnessACP
Kimi CodeACP
MiniMax CodeACP
ZCodeACP
PiACP
Any ACP agent

The problem

Agents now do much of the work. People still carry the context between them, by hand.

Message Claude Code…same brief, again

Message Cursor…same brief, again

Message Codex…same brief, again

01 · One person, many agents

You repeat yourself to every tool

The same background, goals and constraints, pasted into each agent at the start of every session.

Dev A

Bundle pricing

Loyalty rules

Dev B

Port modal

Checkout copy

02 · A team on one agent

Sessions stay invisible

Everyone uses the same tool, but each person's sessions and the decisions in them stay locked on their own laptop.

Why does bundles.ts skip the loyalty check?

The answer is in Dev A's sessionnot shared

03 · A team on different agents

Context moves only by copy and paste

Dev B hits a question about code Dev A's agent wrote. The agent knows the answer, and nobody can ask it.

Delegation

Agents that ask each other for help. Any agent can hand a task to another mid-turn, then wait for the answer or keep working.

Claude Codelead · delegates 3 tasks
Codex--wait
Why does applyLoyalty() run twice?
Answered
Cursor--need figma --write
Match the checkout modal to the Figma frame
Ready to merge
OpenCode--best-of 3 --write
Speed up price calculation
Merged best of 3

Wait
--wait
Block on an answer you need right now, then carry on with it.
Detach
default
Launch the task and keep working. Collect the result when it lands.
Fan out
--harness ×N
Send one task to several agents at once, or to the same agent twice.
Route by need
--need
Pick a ready agent that has the skill, plugin or MCP server the step requires.

Handoff

Switch agents mid-task. Keep everything you've built up.

Starts from the compressed conversation, not a blank slate.
Carries a context pack: decisions, constraints and findings.
Moves the worktree, uncommitted changes included.

Make the cache production readycache

Enforce the 64 MB limit in lru.rs and evict before inserting.

Added a byte-based budget and eviction in src/cache/lru.rs. Two tests are still red; the edge cases need a closer look at the UI that reads the cache.

Continued on Cursor · carried pack vd2e1f2f, 2 summaries, last 2 messages

Message Claude Code… (/ for commands and skills, ! for shell)

Message Cursor… (/ for commands and skills, ! for shell)

Default model

Continue on another harness

A new chat on the harness you pick takes over this Claude Code chat's work. It starts from the compressed conversation instead of a blank slate.

Codexsigned in

Cursorsigned in

OpenCodesigned in

Carrying

pack vd2e1f2f (0.5 KB) · 2 summaries · last 2 messages · state block

The worktree moves to the new chat with its uncommitted changes.

CancelContinue on CodexContinue on Cursor

Collaboration rooms

One room for your team and all of its agents. Edit together, share one board, and never work from stale context.

cacheMake the cache production ready main WORKTRUNKMRPRLive

src/cache/lru.rs 2 editing

12pub fn insert(&mut self, key: String, val: Vec<u8>) {

13 let size = key.len() + val.len();

14 while self.bytes + size > self.limit {Priya

15 self.evict_oldest();

16 }

17 self.bytes += size;

18 self.map.insert(key, val);

19 debug_assert!(self.bytes <= self.limit);Claude Code

20}

Live, together
Everyone in the room sees each cursor as it types, whether it belongs to a person or an agent.
A board for agents
Queue tasks and assign them to any agent. Each one works in its own worktree and moves to review when done.
Context that catches up
When an agent falls behind what the room has decided, send it an update in one click.

pricing/bundles.ts Blame

Claude Code3w1export function applyBundleRules(cart: Cart): Cart {

2 const bundles = groupBundles(cart.items);

Codex3w3 for (const b of bundles) {

AArjun2w4 if (b.loyaltyApplied) continue;

Codex3w5 b.price = b.price * (1 - cart.loyaltyRate);

6 }

PPriyalive7 // loyalty is applied once, upstream in applyLoyalty()

Claude Code3w8 return withTotals(cart, bundles);

9}

10

11function withTotals(cart: Cart, bundles: Bundle[]) {

12 const subtotal = sum(bundles.map((b) => b.price));

13 return { ...cart, bundles, subtotal };

14}

Codexdelegate3 weeks ago

Task “Bundle pricing rules”, delegated by Claude Code

commit 76ffd1d · merged by Arjun

“Loyalty is applied upstream, so bundles can apply it directly.”

Open conversation

● Claude Code 44%● Codex delegate 33%● Arjun 12%● Priya, live 11%

Agent Blame

Know who really wrote each line. Not just whose name is on the commit.

Attributes every line to an agent, harness, model, session or person.
Covers uncommitted and live edits, not only commits.
One click opens the conversation, and the assumption, behind a line.

Developers

Built for the terminal, too. One CLI that any agent, or you, can call.

Agents running in OrkiLab call orkilab agent to find a capable peer, hand off work and collect results. Scripts and CI can use it the same way.

~/checkout-service

❯ orkilab agent delegate --need figma --explain "Match the checkout modal to the Figma frame"

checking ready harnesses for a skill, plugin or MCP server named figma

cursor figma MCP · ready · preferred for UI work

claude-code no match

→ routed to cursor (delegate chat_b81d07e3)

Under the hood

Everything an agent team needs. On your machine, under your control.

Needs review · 1

The cache must stay under 64 MB

Constraint found by Claude Code

Accept AcceptedNow in every agent's contextReject

Shared project knowledge

Agents record findings and constraints as they work. You review each one once, and every agent sees it from then on.

Context window38.5k / 1M

Compaction window38.5k / 80k

Summaries: 1 · 9.2k → 1.3k tokens

Context that stays small

Long sessions are compressed into summaries you can inspect and restore, so handoffs carry the decisions without the noise.

loyalty

pnpm test bundlesfailed
pricing/bundles.tsedit
“Loyalty applies once, before…”reply

One timeline across agents

Every command, output, edit and reply from every agent lands in one searchable timeline that agents can query too.

main +72 −0

orkilab/chat/4f2a9c1e

orkilab/chat/b81d07e3

A worktree for every agent

Agents that edit code each work in their own Worktrunk worktree. Their changes come back as patches you merge, so parallel work never collides.

Runs on your machine

Projects, conversations and worktrees stay local. OrkiLab drives the agent CLIs you already have installed and signed in to.

Open to any agent

Built-in support for the major harnesses, presets for seven ACP agents, and a custom slot for anything else that speaks the Agent Client Protocol.

Coming next

Orki Bots. Something new is joining your agent team.

We're building the next layer of OrkiLab right now. People on the waitlist will see it first.

Get notified Next in line after the beta

Explain it once. Let your agents handle the rest.

We're letting teams into the private beta in small batches. Join the waitlist for early access.