a
LOCAL-FIRST · NO API KEYS · READ-ONLY

A council, on call.

aiki convenes the AI CLIs you already run — Claude Code, Codex, and Gemini — as a panel that genuinely disagrees. They investigate independently, cross-examine, a judge adjudicates the disputes, and you get a decision brief you can audit.

$ npm install -g aiki-cli
Get started → ★ Star on GitHub
C ClaudeCLI detected
C CodexCLI detected
G GeminiCLI detected
262 tests passing No API keys Read-only orchestration Node ≥ 20 · macOS / Linux MIT licensed

The problem

One model has
one blind spot.

On a code review or an "is this idea sound?" call, a single model misses roughly 1 in 4 real problems. Two or three different models — each analysing independently, then cross-examining — catch what any one misses.

aiki orchestrates that locally. No new subscriptions, no copy-pasting between tools by hand — it drives the CLIs and logins you already pay for.

Claude Codex Gemini CAUGHT

The council

Three seats. Real disagreement.

Each provider runs as a local child process with read-only flags. Roles are configurable — pin who judges and who reviews in .aiki/config.json.

C
Claude
judge

Adjudicates disputes and writes the chairman's reasoning. Runs via claude --permission-mode plan.

CLI · claude
C
Codex
verifier

Independent review, then deterministic file:line validation — every finding must point at a real line. Runs via codex --sandbox read-only.

CLI · codex
G
Gemini
analyst

Adversarial analysis from a fresh angle via Antigravity. Runs via agy --sandbox.

CLI · agy

How it works — no APIs

Orchestration, not a hosted service.

01
Spawn CLIs

Your signed-in provider CLIs start as read-only child processes.

02
Blind review

Each model gets the same task and returns structured JSON — no peeking.

03
Cross-examine

Models see each other's findings and argue — consensus vs. dispute map.

04
Judge decides

A judge adjudicates only the disputes — consensus is trusted as-is.

05
Decision brief

Every stage zod-validated, audit trail under .aiki/, rendered as a brief.

Live run

Watch the council deliberate.

aiki — council session

Benchmark · pre-registered · held-out

The council caught every planted bug.
The best single model missed 1 in 4.

Best single model · arm B 0%
33 / 43 seeded bugs · 10 provider calls
aiki council · arm D 0%
43 / 43 seeded bugs · Claude + Codex review, Gemini judges · 44 calls
the recall — at identical precision
0
false positives across 59 adjudicated findings
CaseSeededSingle (B)Council (D)
Honest caveat: the win is a recall win — precision was non-discriminating on this bug-dense set. n = 10 cases, single run per arm — directional, not a p-value. Full method in RESULTS.md. Reproduce: aiki bench code-review --arms B,D --set holdout --yes

Two jobs, done well

Review the code. Stress-test the idea.

/review

Code review

Catch the bugs a single model misses on your working tree or a branch.

Parallel blind review of the diff
Deterministic file:line validation
Mutual adversarial cross-examination
Judge adjudicates only the disputes
$ aiki run code-review --base main
~5 provider calls per run
/idea

Idea stress-testing

Pressure-test a plan before you build it — get a decision brief, not an essay.

Contextual preflight pins your intent
Parallel adversarial analysis
Disagreement map → verifier → judge
Anchored validation plan with kill signals
$ aiki run idea-refinement "..."
~12 provider calls · auto-opens the report

The decision brief

A verdict you can audit.

PROCEED PROCEED WITH CONDITIONS PIVOT STOP
BLUF recommendation

The bottom line up front — proceed, pivot, or stop.

Chairman's reasoning

What decided it, where the models split, whose side the judge took.

Dimension scorecard

Which of 12 angles were examined, contested, or missed.

Assumption audit

Held / failed / unverified — each with a confidence level.

The debate

Who argued what, and who won the point.

Anchored validation plan

Concrete next actions, each with an effort estimate and a kill signal.

Safety model

Trustworthy to point at a real repo.

aiki never uses --dangerously-skip-permissions or any edit/exec mode. It reads, reasons, and writes only its own audit trail.

Read-only orchestration

Every provider runs with its read-only / sandbox flag. No edits, no exec.

No credentials, ever

Never reads credential folders; filters KEY|TOKEN|SECRET from the env before spawning.

Writes only under .aiki/

Nothing else on your disk is touched. Every stage is schema-validated before the next sees it.

The command palette

Type a slash, or just describe an idea.

/mo → filters as you type · Tab completes · ↑/↓ pick · Enter runs

Convene your council.

Three steps to a verdict you can audit — using the CLIs you already have.

# 1 · install
$ npm install -g aiki-cli
# 2 · verify your providers
$ aiki doctor
# 3 · open the home screen
$ aiki
View on npm → Read the docs