Multi-Agent Code Dispatch Dock
A local control plane for parallel coding agents that isolates each workspace, leases contested files, and delivers tested patches in dependency order.
When developers launch several local coding agents at once, the problem quickly shifts from who writes the code to who is changing the same file. They drag Linear or GitHub issues into a dispatch board, specify dependencies, test commands, and the directories each agent may touch. Before an agent starts, it receives its own worktree, environment variables, and disposable test container, while the main branch stays clean.
Before changing a high-conflict file, an agent requests a short-term file lease. If critical areas such as payment modules or configuration files are already occupied, the dispatcher reassigns it to parallelizable testing, documentation, or low-conflict work, or makes it wait for the preceding patch’s result. Once an agent finishes, the system runs the specified tests in its own container and records the code diff, test output, and task context it referenced.
Patches that pass testing enter a merge queue in dependency order. If a later patch depends on an earlier change, it is revalidated against the updated baseline first; failed tasks return to the developer with the terminal state intact. The first release supports only local Git repositories, containerized testing, and file leases. It solves agents overwriting one another, rather than replacing a team’s code review or release permissions.
Why now
As observed on September 8, 2026, Airuncode ranked third in Product Hunt’s new-product feed and explicitly promotes running multiple coding agents locally. S1 Once users begin launching agents in parallel, file contention, interference between test environments, and patch merge order become immediate operational problems.
Target user
Independent developers and small-team leads running two to five local coding agents at once. After breaking down a set of related issues, they want implementation, testing, and documentation to move in parallel. Multiple tasks then touch configuration, type definitions, or shared interfaces, and manually managing worktrees, terminals, and merge order begins to consume attention.
Minimal entry point
Keep core state in local SQLite, recording tasks, dependencies, leases, and run results. Create a separate branch and directory for each task with Git worktrees, an isolation pattern validated by similar tools. S3S4 Use Docker Engine for disposable test containers, with task-level namespaces for ports, caches, and environment variables. Agents request leases through wrapped file-write tools rather than by locking an entire worktree. Support either GitHub Issues or Linear in the first release, not both, to avoid maintaining two sync paths. The merge queue should begin with topological sorting, rebasing, and specified tests, without attempting to judge code quality automatically.
Punching above its weight
The strongest demo is a reproducible conflict: two agents edit the same configuration file, and the dispatcher redirects one to add tests. Open-source the sample repository, run logs, and CLI as a template to reach developers already using Claude Code, Codex, and other local agents. Position the launch around reduced rework, not agent count. After importing their own repository, users can view a file-overlap report for free before deciding whether to enable automatic scheduling.
Competitors & gaps
- AiruncodeGoogle
- Airuncode already offers a local-first environment for running multiple agents. Users bring their own model keys, while code and prompts stay on their machine. It also highlights shared memory, automatic test completion, and repair after failures. S2 That already covers the main entry point of running multiple agents at once. Its public materials focus more on agent capabilities and the testing loop. They do not show a file-lease workflow before edits begin or explain whether tasks are reassigned based on occupied files. The opening for this card is before an agent starts editing: directory permissions, dependencies, and leases inform the scheduling decision. That opening would narrow quickly if Airuncode adds conflict prediction.
- Open OrchestratorGoogle
- Open Orchestrator already creates isolated worktrees for agents, supports multiple coding agents through a unified console, and uses Conflict Guard to detect overlapping file edits in real time. Its merge queue can also suggest ordering and deliver changes one at a time. S3 This closely overlaps with the latter half of this card. The difference is that it primarily monitors edit overlaps after they occur. This concept requires an agent to obtain a short-term lease before writing. When a file is occupied, a task can be redirected to tests or documentation, while dependencies from Linear or GitHub issues also shape the schedule. The difference must materially reduce waiting and rework; otherwise, it is a scheduling feature in an existing tool rather than a standalone product.
How it makes money
A local single-repository setup with basic leases is free. The Pro plan is priced per developer per month and adds multi-repository scheduling, task-system sync, container templates, and historical audit trails. A Team plan adds shared policies, access controls, and centralized execution nodes.
The case against
File leases may mistake healthy parallel work for conflict and leave agents waiting unnecessarily. Agents may also bypass path rules through scripts, generators, or renames. Worktrees isolate files only; databases, ports, caches, and external services can still interfere with one another. Starting a container for every task adds disk usage and wait time. A dependency graph entered manually will quickly drift, while model-generated dependencies can omit items. The more immediate threat is that adjacent tools already offer isolation, conflict detection, and merge queues. S3 If proactive leases do not materially reduce rework, this is better suited as a plugin feature.