← Back to Showcase
Parallel Task Orchestrator · Open Source

claude-fleet

Run hundreds of Claude Code sessions in parallel

403
Tasks Executed
48h
Wall Clock Time
15
Repositories
30
Concurrent Sessions
The Problem

Claude Code is Sequential.
You Pay for a Firehose, Use a Garden Hose.

Without Fleet

1 session at a time

1 session × 12 min/task × 403 tasks

= 80+ hours

Spread across weeks. Context lost between sessions. Human bottleneck on every task handoff.

You have a Max plan. You could run 30 sessions. Instead you run 1 and wait. That is 97% waste.

With Fleet

30 sessions in parallel

30 sessions × 12 min/task × ~14 tasks each

= done in hours

Each session gets its own worktree. No conflicts. No context switching. Results collected automatically.

Same plan. Same rate limits. 30x the throughput. The only thing that changed is orchestration.
Architecture

How It Works

Write tasks → Convert → Run → Monitor

The Workflow

1
Write tasks.md — one heading per task, with context and instructions
2
fleet-convert tasks.md → generates tasks.json
3
fleet tasks.json → spawns parallel Claude sessions in git worktrees
4
fleet-dashboard → live monitoring of all sessions

Config Hierarchy

~/.fleetrc → global defaults
.fleetrc → per-repo overrides
tasks.json → per-task tool permissions

7 CLI Commands

CommandPurpose
fleetRun tasks in parallel Claude sessions
fleet-convertConvert tasks.md to tasks.json
fleet-dashboardLive TUI for monitoring sessions
fleet-reviewReview and merge completed task branches
fleet-statusQuick status check of running fleet
fleet-taskAI-generate task files from descriptions
fleet-cleanupRemove worktrees and temp branches
Key Feature

Slash Commands — Fleet Inside Claude

Stay in your session. No terminal switching. No context loss.

Available Commands

/fleet <tasks>Spawn parallel sessions from within Claude
/fleet-task <desc>AI generates task files from natural language
/fleet-statusMonitor running fleets without leaving
The AI writes better task prompts than you would. Each generated task includes the right file paths, the right context, and the right acceptance criteria.
> /fleet-task "audit payment, booking, and allocation for error handling"
Creates tasks-audit-services.md with 3 focused tasks
> /fleet tasks-audit-services.md
Spawns 3 parallel sessions in background
> /fleet-status
Fleet Status: 1/3 done | 2 running | 0 pending
You never leave your Claude session.
You keep working while fleet runs in the background.
Case Study

NammaYatri: 403 Tasks Across 5 Waves

30 concurrent sessions (3 orchestrators × 10 parallel)

WaveTasksPurposeOutput
1 Audit 225 Analyze every service for error handling, logging, resilience 225 structured audit reports with severity-ranked findings
2 Fix 40 Implement fixes for critical findings from audit wave 40 PRs with code changes, each referencing its audit
3 Verify 41 Validate fixes, check for regressions, confirm patterns 41 verification reports with pass/fail status
4 Critical 22 Deep-dive into highest-severity issues across services 22 detailed analyses with root cause and remediation
5 Observability 13 Add structured logging, metrics, tracing across platform 13 PRs adding observability instrumentation
15
Repositories
48h
Total Wall Clock
3 × 10
Orchestrators × Parallel
Pipeline

The Overnight Pipeline

Friday evening → Saturday morning: read the reports

Audit
225 tasks
Analyze every service
Output: structured reports
Fix
40 tasks
Implement changes
Output: PRs with code
Verify
41 tasks
Validate & check regressions
Output: pass/fail reports

Multiple Orchestrators

Run 3 fleet instances simultaneously, each managing 10 parallel sessions. One per repo group. They share nothing — each gets its own worktrees, its own branches, its own output directory.

Pipeline Chaining

Each wave's output feeds the next wave's input. Audit reports become fix task context. Fix branches become verify targets. fleet-convert accepts previous results as context for the next wave.

Start the pipeline Friday evening. By Saturday morning, you have 225 audit reports, 40 PRs, and 41 verification results waiting for review.
Summary

Zero Infrastructure. Any Codebase. 200 Lines of Bash.

Zero Infrastructure

No servers. No containers. No cloud accounts. Just bash scripts, git worktrees, and your existing Claude subscription.

Any Codebase

Works with any language, any repo size, any CI system. If Claude Code can work on it, fleet can parallelize it.

Granular Control

Per-task tool permissions. Read-only audits get no write tools. Fix tasks get Bash + Edit. You control exactly what each session can do.

~200
Lines of Bash
7
CLI Commands
3
Slash Commands
0
Dependencies

github.com/nammayatri/claude-fleet

"The gap between 'Claude Code can do this' and 'we actually did this across our entire platform' is just a task file and a fleet command."

1 / 7