the agent harness

Run agents like a team,
not like tabs.

Groovy turns Claude Code and Codex into a workforce. One orchestrator you talk to, named agents that do the work — in your repos, on your machine, with your keys.

5-day free trial · your API keys · macOS & Windows

groovy — harness connector online
FixterClaude Codeopus-4.7
idle — awaiting dispatch
ReviewerCodexgpt-5.6-sol
idle — awaiting dispatch
Tasks

Coding agents got good. Managing them didn't.You're alt-tabbing between terminals, re-pasting context, babysitting runs. That's not orchestration — that's a second job.

// how it works

Hire. Dispatch. Ship.

01

Hire

new agent → Fixter · Claude Code · ~/repos/api

An agent is a real coding harness — Claude Code or Codex — bound to a repo on your machine. Give it a name, a workspace, a model, and the skills it should know.

02

Dispatch

@Fixter fix the flaky auth test

Talk to one orchestrator. @mention an agent for precision, or describe the outcome and let it route, sequence, and hand context between agents for you.

03

Ship

✓ Fixter finished · 2 files changed · +41 −12

Watch every agent work live in the grid. Risky work waits for your approval. Done work pings your phone. Transcripts, diffs, and costs are all on the record.

// model economics

Fable 5 plans.
Haiku executes.
You keep the difference.

Every role gets its own model — the orchestrator brain, each agent, even your daily heartbeat digest. Pick from the frontier catalog or paste any model id. When the bill comes, ask the orchestrator to read it and it will tell you which agents can drop to a cheaper model without dropping quality.

claude-fable-5 · opus-4.7 · sonnet-4.6 · haiku-4.5
gpt-5.6-sol · gpt-5.6-terra · gpt-5.6-luna · gpt-5.5 · any-model-id

OrchestratorFable 5
Fixter · Claude CodeOpus 4.7
Reviewer · CodexGPT-5.6 Sol
Docsmith · Claude CodeHaiku 4.5
Heartbeat digestGPT-5.6 Luna
// the harness

Everything between “I have agents” and “they run my backlog”

Plans are files, not chat scrolls

Flip on plan mode and an agent drafts read-only. Approve it and the plan lands in the repo — where Claude Code, Codex, and your teammates can all read it.

.claude/plans/2026-07-10-auth-fix.md

Skills from your own repo

Markdown playbooks, synced from git, assigned per agent. Claude Code agents get CLAUDE.md context, Codex agents get AGENTS.md — automatically materialized on device.

skills/deploy-checklist.md → Fixter

Agents on cron

Tell the orchestrator when and who: nightly test triage, weekly dependency bumps, or a 7:30 report. Each run uses that agent's harness, workspace, model, and skills while its connector machine is awake and online.

daily @ 07:30 → Scout: "triage new issues"

WhatsApp is your remote

Dispatch from your phone, get pinged when work lands, and approve or reject with a reply. The harness doesn't stop because you stood up.

"approve 3f2a91" → running

Context moves between agents

One command summarizes what an agent learned and briefs another — findings from Scout become marching orders for Fixter. No copy-paste archaeology.

transfer_context Scout → Fixter

Spend, by agent, with receipts

Every token is attributed to an agent, a model, and an outcome. Ask the orchestrator to analyze it and it recommends cheaper model mixes that won't cost you quality.

usage_report --days 30

Plus data integrations (Gmail, Calendar, ads platforms, Postgres, Firecrawl…) your agents can pull from — see the catalog.

// trust model

Powerful agents need a short leash.

Groovy holds it: local execution, your credentials, human approval on anything that can hurt.

Runs on your machine

Agents execute in your repos through a local connector. Your code never transits our servers.

Your keys, your bill

Anthropic, OpenAI, Google, Azure, Bedrock — bring your own credentials. Groovy adds no token markup.

Approval where it matters

Destructive work waits for a human. Plans wait for review. Everything else just ships.

Source-available

A delayed public mirror keeps the code inspectable. Trust is something you can read.

// groovy desktop

One download.
Zero pairing codes.

Groovy Desktop bundles the harness and the local connector in a single signed app. Install, sign in, and your machine links itself. Agents keep running when the window closes; updates are one click, ChatGPT-style.

  • Connector managed for you — starts, restarts, updates
  • Keeps schedules and heartbeats alive in the background
  • Signed & notarized DMG for Apple Silicon
Groovy.dmg
macOS 12+ · Apple Silicon
Download Groovy Desktop

Prefer headless? The standalone connector still works everywhere.

// pricing

A license, not a meter.

Every account starts with a 5-day free trial. After that it's a flat yearly license — your model spend goes straight to your providers, with no Groovy markup in the middle.

Groovy Personal

For personal projects

$49.99per year
  • One individual, personal non-commercial use.
  • Two activated devices.
  • Current downloads, updates, and source snapshots while active.
Groovy Enterprise

For companies

Contact salesannual license
  • Commercial use, self-hosting, and internal modification rights.
  • Source access, support, security, and contract terms.
  • Fallback rights to the last paid version if you do not renew.
Default billing

No token tax

$0Groovy token markup
  • Connect your own OpenAI, Anthropic, Google, Azure, Bedrock, Groq, or Mistral keys.
  • You control provider usage and provider billing.
  • Groovy usage analytics are separate from Groovy licensing.
Approved partners

Reseller billing

By agreementauthorized only
  • Usage-based billing stays off unless the license explicitly enables it.
  • Approved resellers can meter customer usage and export billing data.
  • Normal personal and enterprise customers never see reseller controls.

Your agents are ready.
Give them a manager.

Five days free. Your keys, your machine, your call.

Personal license $49.99/yr No token markupmacOS 12+ (Apple Silicon) · Windows 10+ (x64)