Skip to main content
Glama

What it adds

Irrigator runs coding agents in parallel while keeping results, retries and spending evidence on disk. Use the CLI directly or through Claude Code, Codex and Antigravity plugins. Inspect the work and resume from your terminal.

Capability

What it gives you

Durable execution

Saved plans, attempts and accepted artifacts that survive interruptions.

Coding cells

One task per git worktree, with branches and objective checks to review.

Bounded delegation

Explicit task graphs, worker concurrency and provider limits.

Recovery decisions

Uncertain external work stays visible and is never silently replayed.

Terminal review

status --watch, graph, review tui and a read-only MCP view.

Related MCP server: coding-agent-mcp

Try it without credentials

Requires Go 1.26.7+, make and a POSIX shell. Development and CI pin Go 1.27.2. From a source checkout:

git clone https://github.com/grainulation/irrigator.git
cd irrigator
make build
./irrigator doctor
./irrigator demo

The demo uses mock workers, needs no credentials and makes no model calls. Each invocation creates a fresh run directory and prints inspection commands. For your own plan, follow the first-run guide.

To install from this checkout, run make install PREFIX="$HOME/.local" and add $HOME/.local/bin to PATH. Installation copies the CLI; agent plugins and provider tools have their own setup. CLI archives and their installation checks are described in runtime releases.

Homebrew users can install the stable CLI from this repository's tap:

brew tap grainulation/irrigator https://github.com/grainulation/irrigator
brew install grainulation/irrigator/irrigator
irrigator demo

The stable formula installs a pinned release archive and does not need Go. For a development source build, use brew install --HEAD grainulation/irrigator/irrigator. See the Homebrew guide for supported platforms, upgrades and checks.

Ask your agent to start

Give Claude Code, Codex or Antigravity this prompt from the checkout:

Read README.md and AGENTS.md in the Irrigator repository. Build ./irrigator with make build,
run ./irrigator agents and ./irrigator demo, and report the exit codes and final status lines.
Tell me which routes below this machine can run. Ask before any model-backed or paid run.

Route

What it does

Needs

Cost and limits

1. Mock demo and starter plan

demo, init, plan, apply --backend mock

Local CLI

No model calls or API spend

2. Local models

Managed tasks using Ollama

CLI, Ollama and a local model

Local compute; no API spend

3. Managed CLI profiles

Turn/helper graphs via claude:MODEL, codex:MODEL, agy:MODEL

CLI and authenticated provider

Cost tracking by default; explicit enforced-budget profiles available

4. Coding cells

irrigator code, with branches to review

CLI, git repo and an agent CLI

Per-cell timeout; Claude USD cap; Codex/agy cost unknown, no USD cap

Real providers use your account or local compute. New managed plans track estimates without blocking when costs are unavailable; tracking thresholds warn. Enforced budgets control admission and reservations and can still be overrun by a provider. Unknown cost never means free execution. See cost tracking and enforced budgets.

Put it to work

Start with a plan, validate it, then choose a backend explicitly:

./irrigator init --out plan.json
./irrigator plan plan.json
./irrigator apply plan.json --run runs/first --backend mock
./irrigator status runs/first

Use new plan filenames and run directories. For repository changes, coding cells accept --task ID=PROMPT and --verify ID=COMMAND; results land on local branches for review. A verification timeout or exit 77 leaves the result uncertain, as does a successful check that changes the worktree or Git identity. See coding cells for setup, spending limits and recovery.

For graphs that grow during execution, choose recursive spawning or managed turns and helpers. These have separate execution rules; worker concurrency is independent of graph size.

Plugins

Install the irrigator runtime separately, then follow agent setup for your host:

The plugins share Markdown skills and the local CLI under the host's normal execution permissions. Host integration covers installation and discovery checks; static package checks alone do not establish that a host can load or invoke a plugin.

Explore the docs

Guide

Start here for…

First run

Your own plan, provider selection and inspection.

Coding cells

Parallel repository tasks and verification commands.

Workflow composition

Reusable graphs and artifact contracts.

Terminal workspace

Swarm dashboard and local review.

Recovery

Interrupted calls, retries and execution authority.

Architecture

Source layout, package layers and durable state.

Compatibility

The supported 1.x interfaces and saved-run formats.

Documentation index

Accounting, providers, command references and more.

Verification and limits

The CLI supports macOS and Linux. Windows users run the Linux build inside WSL 2; native Windows is planned. Platform support records the tested configurations.

Worktrees keep concurrent checkouts apart; they are not a host sandbox. Approvals record a decision bound to a specific head and never merge changes. See SECURITY.md for the local trust model.

CI runs builds, mock workflows, race tests and package checks. Native host and live-provider acceptance are separate. Run the development gates with:

make acceptance
make lint

The verification matrix details their scope, including integration tests. Contributing covers setup and changes; choosing a tool explains the tradeoffs. See the changelog for releases and the release checklist for artifact promotion.


Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables AI coding agents to spawn, parallelize, and synchronize multi-agent programming swarms through a stdio MCP interface. Provides tools for DAG-scheduled task dispatch, isolated Git worktree execution, atomic file locking, shared blackboard memory, and clean merging of completed agent work back to the main repository.
      1
      MIT