kiloloop · the operator loop

AI-agent fleet work,
made reliable.

An open protocol and operating loop for Claude, Codex, and custom agents — with a local-first cockpit when you want one.

See every handoff. Review every change. Persist every lesson. Keep execution local.

See the loop →Read the protocol See Explorer

no daemons · no central server · just files

OACP Explorer showing the agent fleet and today's message feed

Explorer, the optional cockpit · the record is the protocol's, with or without it

The loop08 STEPS · 01 RECORD

One pass around the loop, and the work is accountable.

Each step hands the next one a file. Nothing runs in a vendor's memory; every transition is written down, signable, and yours to keep.

The loop doesn't just run. It leaves receipts.

Signed

Messages can carry an Ed25519 signature, and enforce mode rejects anything unsigned or unverified at intake. Every autonomy decision gets an audit record. Not a log you trust — a record you check.

Verifiable

Verification runs against the files, not against us. Nothing in the chain needs kiloloop to be online, or to exist.

Portable

Claude, Codex, or an agent you wrote last night — the record reads the same. Org-owned, vendor-neutral, cross-runtime.

ZCode vs Codex vs Claude Code, compared →

Step 02Dispatch

# inbox/codex/
├── msg_a8f1c2.yml review_request
├── msg_b2e445.yml review_lgtm
└── msg_c3f0a1.yml task_request
sig ed25519 ✓ · thread #42 · 3 messages

OACPprotocol · v0.5

The coordination layer.

File-based, zero-server, Apache-2.0. v0.4 adds Ed25519 message signing with enforce-mode verification, and records every autonomy decision — the receipts the rest of the loop is built on.

oacp.dev →

Step 08Observe

oacp explorer
PROJECTS
· sites
· oacp
· hq
codex ▸ iris
msg_a8f1 handoff
msg_b2e4 review
msg_c3f0 dispatch ●

Explorercockpit · mac

Watch the whole fleet.

The cockpit — optional. A desktop window onto live fleet state: projects, agents, inboxes, threads, read from your own machine. The loop runs without it; Explorer is how you watch it.

/explorer →releasing soon

Step 01Estimate

$ agent-estimate --issues 11,12,14
Task            PERT      Expected
Auth flow       25/50/90  57.8m
Tests           12/23/40  24.0m
! exceeds Opus p80

agent-estimateplanning · v0.8.0

Know before you build.

Planning. PERT for agent work, METR thresholds, wave plans — as a CLI and as a GitHub Action on the Marketplace.

/agent-estimate →on marketplace

Steps 05 · 07Debrief · Sync

$ agent-memory recall
## 2026-09-07
- Use SQLite for the cache. Why: no daemon
  (codex, 2026-09-07T01:53:42Z)
4 project files · 3 org files · 8,000 chars

agent-memorycli · v0.1.3

Memory that survives the session.

Cross-session memory for coding agents. Plain Markdown files, git-native, no server. Claude Code and Codex share one store; recall reads a fixed file set, capture appends one dated decision.

/agent-memory →new

And the rest of the loop

oacp-skillsExecuteThe operating-loop skills for OACP runtimes — check-inbox, doctor, review-loop, self-improve, wrap-up. Documented at oacp.dev.kiloloop-skillsExecuteAgent-discipline skills, runtime-agnostic. Self-contained, versioned, fixture-tested.memory-lintImproveMemory hygiene, mechanized. Deterministic and flags-only: it reports findings, it never writes.brainstormReviewMulti-model decisions. One question to three runtimes; synthesize the union, keep the disagreement.

coordinator intelligenceAll stepsRuns the loop on your behalf — estimates, dispatches, chases the review. In development.

Tools & Skillsall tools & skills →

Small tools. Sharp edges.
Work you can check.

No auto-fix. No magic. Just findings. Every tool in this line is deterministic, versioned, and fixture-tested — the same discipline the loop asks of your agents, applied to the tools themselves.

$ memory-lint --config memory-lint.yaml
MEMORY.md:12  broken-link  [[archive/old]]
tasks.md:4    stale        180d
index.md:—    drift        2 missing
3 findings · 0 fixed — flags only

memory-lintcli · v0.1.0

Memory rots. Catch it.

A deterministic, flags-only linter for Markdown memory and documentation corpora. It reads target files and reports findings; it has no write or auto-fix path.

pip install memory-lintnew

skills/memory-lint/
├── SKILL.md
├── scripts/memory_lint_wrapper.py
└── fixtures/ 8 passing
pinned: memory-lint==0.1.0

kiloloop-skillsskills · v0.4.3

Discipline, packaged.

Agent-discipline skills for coding-agent runtimes (Claude Code, Codex): self-contained, versioned, fixture-tested. Each skill carries its own instructions, scripts, and falsifiable fixtures.

github →new

Building on OACP? The runtime skills live at oacp.dev ↗

The wedge

Everyone is shipping a fleet view.
Nobody is shipping the fleet record.

What the dashboards give you

A view over someone else's cloud.Agents you can only watch inside one runtime.History that ends when the contract does.

What the loop leaves behind

Signed messages you can verify without us.One record that reads the same from Claude, Codex, or your own agent.Files on your disk, in your org's repo, under your key.

Portable. Auditable. Yours.

The record survives the vendor.

Not another agent OS. The reliability layer under your agent fleet.

Explore OACP →See the cockpit

dogfooded across 600+ validated agent dispatches · as of Sep 2026

Founder

Built by a solo founder, in public.

The same agent fleet that runs kiloloop produced the code, the docs, and the estimates you see here.

hello@kiloloop.com

AI-agent fleet work,made reliable.