granary

Granary

Shared memory and coordination for AI coding agents. Single binary, local-first, no cloud.

# Install
curl -sSfL https://raw.githubusercontent.com/speakeasy-api/granary/main/scripts/install.sh | sh

# Init + plan + watch
granary init
claude -p "use granary to plan: Migrate endpoints to v2" & granary summary --watch

Getting Started

Install

macOS / Linux:

curl -sSfL https://raw.githubusercontent.com/speakeasy-api/granary/main/scripts/install.sh | sh

Windows (PowerShell):

irm https://raw.githubusercontent.com/speakeasy-api/granary/main/scripts/install.ps1 | iex

From source (requires Rust):

cargo install --git https://github.com/speakeasy-api/granary.git

LLM-First workflow

granary init
granary plan "User authentication"
granary next
granary work <task-id>
granary task <task-id> done
granary summary --format prompt

plan creates a project and breaks it into tasks. work claims a task and gives your agent full context. summary produces an LLM-optimized handoff of everything that happened.

Runners & Workers

Runners are reusable command configs. Workers subscribe to events and spawn runners automatically. This is the core automation loop — plan work, then let workers drive execution.

1. Configure a runner

granary config runners add claude-implementer \
  --command "claude" \
  --arg "-p" \
  --arg "$(granary work start {event.entity_id})" \
  --arg "--allowedTools" \
  --arg "Bash,Read,Write,Edit,Glob,Grep" \
  --concurrency 3

2. Start a worker

granary worker start --runner claude-implementer --on task.next

Now every time a task is available, a Claude Code session is spawned with the task context piped in. Add filters to narrow scope:

granary worker start --runner claude-implementer --on task.next --filter "priority=P0"

3. Monitor

granary workers              # List active workers
granary runs                 # List runner executions
granary runs --watch         # Live-updating view
granary run <run-id> logs    # Inspect a specific run

Runner args support {event.entity_id}, {id}, {title}, {project_id}, and other event payload fields. See docs/workers.md for the full reference on events, filters, template substitution, retry behavior, and concurrency control.

Why Granary?

Agents don't coordinate well on their own. Without shared infrastructure they lose context between sessions, duplicate work, and create silent conflicts.

• Session-centric context — explicit "what's in context" for each agent run, so nothing is lost between handoffs
• Lossless planning — agents can clear their working context freely; granary persists decisions and progress for the next agent
• Concurrency safety — task claiming with leases prevents multiple agents from colliding on the same work
• LLM-native commands — plan, work, and initiate bundle multiple operations into single calls, reducing tool invocations
• Event-driven automation — workers react to state changes and spawn agent sessions without human intervention

CLI Workflows

Planning

granary plan "Audit service"
granary initiate "Q1 Platform Migration"

Task management

granary project <project-id> tasks create "Implement login" --priority P0
granary next
granary start <task-id>
granary search "auth"

Context & handoffs

granary context --format prompt --token-budget 2000
granary handoff --to "Code Review Agent" --tasks task-1,task-2
granary summary

Output Formats

Every command supports multiple output formats:

granary tasks                    # Human-readable table
granary tasks --json             # JSON for parsing
granary tasks --format prompt    # Optimized for LLM context
granary tasks --format yaml      # YAML
granary tasks --format md        # Markdown

Watch Mode

granary tasks --watch
granary workers --watch --interval 5
granary runs --watch --status running

Supported commands: tasks, projects, workers, runs, sessions, initiatives, search, summary.

Key Concepts

Concept	Description
Workspace	A directory (typically a repo) containing .granary/
Project	Long-lived initiative with tasks and steering references
Task	Unit of work with status, priority, dependencies, and claiming
Runner	A reusable command configuration (stored in ~/.granary/config.toml)
Worker	A process that subscribes to events and spawns runners
Session	Container for "what's in context" for an agent run

Commands