New v3.9 · parallel reviewer agents For engineering leaders

Parallel AI agents
that ship reviewed pull requests.

aiorch decomposes a single task across multiple coding agents, makes them review each other's work in rounds, resolves conflicts, and delivers a single GitHub PR ready for human review. Runs entirely on your infrastructure. Your model keys. Your code. Zero markup.

Self-hosted, no outbound data BYOK — Claude, OpenAI, Kimi, Codex, Ollama Install in 5 minutes via Docker
session · feat/payment-refactor 4 agents · round 2/3 · $0.71 spent
01 · Task
"Refactor payment module with retry logic and tests"
Decomposed → 4 subtasks
02 · Parallel agents
payment-core claude-sonnet
retry-backoff gpt-5
tests + coverage claude-sonnet
api-docs kimi-k2
03 · Review rounds
reviewer critique · revise
round 1 · 1 revision
round 2 · approved
merge · conflicts resolved
04 · Delivery
Pull Request #847
38/38 tests passed
events.jsonl local
— Model-agnostic · Bring your own key
Claude OpenAI Codex CLI Kimi Ollama · local GitHub
5min
From docker pull to your first orchestrated session.
N
Parallel agents per task — set by your budget, not ours.
0%
Token markup. You pay providers directly at list price.
1
Reviewed pull request per session — ready to merge.
Why aiorch

Most coding tools give you code.
We give you a reviewed PR.

The gap between "AI wrote this" and "a human should merge this" is where senior engineering time disappears. aiorch closes it — by making agents review each other before you ever see the diff.

Single-agent tools

Generates a diff. Hands it to you.

You still need to read it, run it, reason about edge cases, reject half of it, ask for revisions, and open the PR yourself. The review burden stays with the human.

aiorch

Generates, reviews, revises, merges — then opens the PR.

A dedicated reviewer agent critiques each agent's output, sends work back for revision, and only approves once the code passes. You review reviewed code, not draft code.

The code was already reviewed before you saw it. That is the product.

Cross-provider routing

Structurally impossible for
Claude Code or Codex.

Claude Code can only orchestrate Claude models. Codex can only orchestrate OpenAI models. By design, neither can give you cross-brand routing. This isn't a feature gap they'll close — it's a commercial conflict. Anthropic will never natively support GPT-5. OpenAI will never natively support Claude. We will.

One session, any model
Hard tasks
Opus / GPT-5.5
Medium tasks
Sonnet / GPT-5.4
Boilerplate
Haiku / Ollama local

All in a single session. Automatic. Per-agent.

Team-level output, individual cost
Claude Code Teams 3–4×
more tokens for the same work vs individual sessions
aiorch
multi-agent collaboration at individual session cost

aiorch delivers team-style multi-agent orchestration — parallel coding, adversarial review, automatic merge — at the inference cost of individual sessions. No inflated team pricing.

Architecture

Why autonomous works here.
Three review layers, not one.

Earlier autonomous coding tools failed because one agent's mistakes went unchecked until the user saw a broken PR. aiorch replaces human-in-the-loop intervention with adversarial agents at three independent layers — the same problem, a different solution.

01 Reviewer agent in-loop

A separate agent evaluates each coder's output against the original task specification, requests revisions, and only approves once the work passes. Coders never merge their own code. Most issues are caught and corrected before they leave the agent's branch.

02 Independent external auditor

A second-pass reviewer with no shared context audits the merged result, identifies issues the in-loop reviewer missed, and performs final cleanup. Independent context is what makes the audit adversarial rather than confirmatory.

03 Integration verification

Tests run across all merged branches. Conflicts are detected and resolved automatically. The PR only opens if every check passes. If anything fails, the session does not produce a PR — it surfaces the failure for review.

Three checkpoints between the prompt and the merge button. The user sees only the final result.

Pipeline orchestration

Multi-phase work,
not just single tasks.

Real engineering work isn't one prompt. It's a sequence of dependent phases — refactor the schema, migrate the callers, update the tests, deprecate the old API. aiorch chains sessions into pipelines that execute sequentially, with each phase handing off cleanly to the next.

pipeline · dynamic-orchestration-redesign 11 phases · 61 agents · all merged
Phase 1 orchestrator-executor sequential merged 7 agents
Phase 2 test-client sequential merged 5 agents
Phase 3 flip-default sequential merged 4 agents
Phase 4 delete-old sequential merged 7 agents
Phase 5 turn-overlap-pause sequential merged 6 agents
Phase 6 stt-time sequential merged 5 agents
Phase 7 caller-profiles sequential merged 5 agents
Phase 8 ssml-rate sequential merged 6 agents
Phase 9 sentiment sequential merged 5 agents
Phase 10 backchannels sequential merged 6 agents
Phase 11 speculative-stretch sequential merged 5 agents

The pipeline above is a real architectural refactor delivered by aiorch — 11 sequential phases, 61 agents, all merged without manual intervention. Pipelines are how aiorch handles work that would normally span multiple sprints.

Live session

One task. Four agents.
A reviewed pull request.

The session view below shows a real orchestration run. Four agents working in parallel, a reviewer requesting revisions, and the final PR — all streamed token-by-token in your browser.

aiorch — localhost:1230/session/sess_8f2a1c LIVE
[session] payment-core refactor · 4 agents · round 2/3
12:26:32 [planner] decomposed task into 4 agents (3 dev + 1 docs)
12:26:33 [session] smart-routing: a01 → claude-sonnet · a02 → gpt-5.5 · a03 → claude-sonnet · a04 → kimi
12:26:41 [a01] tool: write_file(src/services/payment.ts) +142 / -38
12:27:02 [a02] tool: write_file(src/retry/backoff.ts) +86 / -0
12:27:18 [a03] tool: run_tests(pytest tests/test_retry.py)
12:27:22 ✓ 14 passed in 4.2s · coverage 94.2% (was 78.1%)
12:28:04 [reviewer] round 1/3 · a02 revision requested — edge-case handling
12:28:47 [a02] revised · null-safety + jitter bounds
12:29:12 [reviewer] round 2/3 · all agents approved
12:29:18 [session] merging branches → integration/payment-refactor
12:29:21 ◆ conflict resolved: src/services/payment.ts:142
12:29:34 ✓ integration tests · 38/38 passed
12:29:41 ⬆ PR opened · github.com/acme/platform/pull/847
12:29:41 ✓ session complete · 4 agents · 3m 09s · $0.84 (your keys)
Inside aiorch

An operator console,
not a chat box.

Every running session exposes the full agent fleet, live logs, model spend, and a complete event timeline. Built for engineers who merge the PR — not for demos.

localhost:1230/ aiorch.
Operator control
Sessions & pipelines
Live
Active
2 / 17
Pipelines
1 / 3
Agents
8
Today
$1.84
sess_8f2a1c Reviewing Refactor payment module with exponential backoff retry logic…
sess_a04b29 Merged Migrate REST v2 endpoints to GraphQL schema…
sess_d3e104 Failed Add structured audit logging with Loki exporter…

Dashboard — everything at a glance

/ · landing view
localhost:1230/session/sess_8f2a1c aiorch.
← Session Reviewing · round 2/3
payment-core refactor
Agent Model State Spend
a01 payment-core claude-sonnet-4 Reviewing $0.24
a02 retry-backoff gpt-5.5 Approved $0.18
a03 tests claude-sonnet-4 Running $0.21
a04 api-docs kimi-k2 Approved $0.08

Session detail — fleet + live logs

/session/<id>
Local-first architecture

Your code never leaves your infrastructure.

aiorch is a docker image you run on your own machine, VM, or build server. The orchestrator, agents, git worktrees, and review loop all execute locally. The only traffic that ever leaves the container is the model API call — to the provider whose key you configured.

Zero telemetry on your code

No aiorch server ever sees your prompts, diffs, code, or metrics. The only outbound call beyond your configured model APIs is a one-time license check.

You control the egress

Configure which provider endpoints the container can reach. Point at self-hosted inference, OpenAI-compatible gateways, or local Ollama — the orchestrator doesn't care.

Runs anywhere docker runs

Laptop, bastion host, CI runner, on-prem cluster. Same image, same behavior. No SaaS dependency to audit, no data-processing agreement to sign.

Orchestrator
local
Agents × N
local
Reviewer
local
Git worktrees
local
events.jsonl
local · append-only
Cost ledger
local
your source code
stays inside the container
model API call only
configured provider
claude · openai · kimi · codex · ollama
Bring your own key

Your provider contracts.
Your list price.

We don't resell tokens. We don't inflate margins on inference. Plug in the enterprise agreements you already have — or route agents to your own self-hosted models. Every session ends with an itemized cost report, agent by agent.

$0
platform margin on tokens
100%
spend visibility per session
Anthropic Cloud
claude-opus-4 · claude-sonnet-4
claude-haiku-4
OpenAI Cloud
gpt-5.5 · gpt-5.4 · gpt-5
o4-mini · o3 reasoning
Moonshot Cloud
kimi-for-coding
Codex CLI Cloud
codex-default
OpenAI-backed agent
Ollama Self-host
qwen-coder · deepseek
any tool-capable model
OpenAI-compat Self-host
vLLM · TGI · LM Studio
bring your own gateway
How it works

Five stages, one command, one PR.

aiorch runs a deterministic orchestration pipeline. Each stage is logged and replayable — so you can walk back through any session end-to-end.

01

Decompose

Your task is parsed into isolated subtasks with explicit scope and acceptance criteria.

02

Fan out

Each subtask gets its own agent, git worktree, and model. Agents run in parallel, in isolation.

03

Review

A dedicated reviewer agent critiques each output, demanding revisions until the work meets spec.

04

Merge

Branches are merged into an integration branch. Conflicts are resolved programmatically or escalated.

05

Deliver

A single GitHub PR lands with summary, diff stats, agent breakdown, and complete event log.

Engineer impact

What one engineer actually saves.

Based on internal benchmarks across 100+ sessions at TravelCalls.ai — refactors, test backfill, and feature scaffolding compared to manual workflows. Your numbers will vary by codebase shape and task mix.

Per routine task
Before
With aiorch
Senior-engineer time
~4 hrs
~20 min
Review cycles to merge
2–4 rounds
1 round
Inference cost
$0.50–$2 typical
Typical annual impact per engineer
5–10× faster on routine tasks

Based on internal benchmarks across 100+ aiorch sessions on a production Rust + React codebase. Inference billed at provider list price; aiorch licence purchased separately.

What engineers see first

01

Senior time back on senior work

Routine refactors, test backfill, and scaffolding stop consuming your hours. You go back to architecture, design, and review.

02

PRs ready by morning

Kick off a session before you log off. By standup, the PR is open, reviewed, and waiting on your approval — not your authorship.

03

Visible, bounded inference spend

Every session has a hard budget cap and an itemized cost ledger. Real provider costs, no markup, no vendor surprise invoices.

Audit logs

Every agent action, on disk.

Every session writes a structured event log to your local file system. Know exactly which agent wrote which line, which model was called, and what it cost — replayable from your own machine.

data/debug.log · append-only JSONL local · structured
02:14:06.112 orchestrator task decomposed (4 subtasks)
02:14:48.201 agent-1 PaymentService.ts · patch applied · 214 loc $0.19
02:15:33.408 reviewer agent-2 revision · missing timeout bounds $0.04
02:16:02.115 agent-2 ExponentialBackoff.ts · revised · 88 loc $0.12
02:16:51.040 agent-3 38 tests authored · coverage 94.2% $0.31
02:17:29.778 reviewer round 2 · all agents approved $0.06
02:17:52.002 orchestrator PR #847 opened · integration/payment-refactor $0.12

Structured debug log

Every agent spawn, review round, merge, and completion is appended to an append-only JSONL log on disk. One line per event, grep-friendly. One-click diagnostic export bundles the full session into a single archive.

Real token costs, per agent

Cost is captured from the provider's API response where available, or calculated from published pricing for CLI-based providers. Roll up by agent, by session, or by pipeline.

Activity timeline in the UI

The session detail page renders the full agent timeline: who started when, what they touched, when the reviewer signed off. Same data as the JSONL — just rendered.

One-click diagnostic export

Export a complete bundle for any session or pipeline — events, prompts, diffs, model responses, costs — as a single archive. Drop it in a bug report or keep it for post-mortem.

Comparison

Where aiorch fits in the landscape.

Tools overlap, but they solve different problems. aiorch isn't competing for your IDE — it's competing for the review-and-merge loop.

IDE agents
Cloud autonomous
aiorch
Delivers a pull request
variessome do, some IDE-only
yes
yesreviewed by default
Multi-agent review loop
partialsingle-pass or add-on
partialsingle reviewer
nativemulti-round enforced
Runs on your infrastructure
partialagent may run cloud-side
cloud onlyor VPC at enterprise tier
docker, localfully self-hosted
BYO model / provider keys
limitedown ecosystem only
bundled inference
any provider+ local Ollama
Token markup
seat + creditsvaries by vendor
compute-basedbundled pricing
zeroprovider list price
Cross-provider routing
single vendor
single vendor
any mixper agent, per session
Built in production

aiorch was built while shipping a Rust + React production AI platform.

We needed an orchestration tool that could deliver multi-phase refactors without supervision while we focused on customer-facing work. Nothing on the market did that the way we needed, so we built aiorch. Every feature on this page exists because we used it ourselves first.

— internal use case, not a paid endorsement

Pricing

Per seat. Your tokens.

aiorch is a Docker container that runs on your infrastructure. You pay per seat for the orchestrator. All inference is billed by your model provider at their list price, directly to you.

Free trial
Run aiorch for 14 days with full features. No card, no commitment.
Free14 days
all features · no card required
Install now
  • Full feature access
  • BYOK · all providers
  • Unlimited sessions & agents
  • Converts to paid after 14 days

Named-user licenses. Each seat is bound to one email and can be installed on unlimited devices by that user. No device limits, no concurrent session restrictions. See what gets logged →

Roadmap

What's coming, what isn't.

We list everything below — including what's not yet built. aiorch is local-first by design; if a feature would need a hosted service, we say so.

In progress

Signed event log

Hash-chained, sha-256-signed events.jsonl for tamper-evident replay and post-mortem. Today the log is plain JSONL on your file system.

target · next minor
Planned

Policy & model allowlists

Declarative YAML to constrain which models, repos, and branches a session may touch — plus per-session spend caps enforced at the orchestrator. Useful even for a solo user; not yet shipped.

target · Q3
Considering

Multi-user / shared history

SSO, role-based access, and a shared session history would require a hosted backend. aiorch is single-user by design today; we'd only build this if there's clear demand for a team edition.

no commitment

Wake up to a PR already reviewed.

Install in 5 minutes. Run it against a real branch tonight. Free for 14 days. No card, no phone-home, no tokens to pre-buy.

Start free trial