Mycelium

⚡ Quickstart → Install Mycelium

A task manager built for coding agents. Mycelium gives an AI agent (Claude Code, Cursor, Aider, or your own) a durable, git-trackable place to plan and remember work across sessions. State lives in a SQLite file inside your repo, so when an agent's context is cleared, a session ends, or a teammate pulls the branch three days later, the plan is still there: the same epics, tasks, dependencies, and open follow-ups. The agent reconstructs where it left off from myc list, not from your memory.

It is a one-shot, scriptable CLI with JSON output and a zero-config data model, which is exactly what an agent needs to drive it reliably. Humans get the same tool (plus an optional desktop GUI, MycUI).

A real Claude Code session driving myc to plan work in a repo.

Why agents

• State that outlives the session: the plan is a file in the repo (.mycelium/mycelium.db), not the agent's context window. Clear the context, start a new session, or hand the branch to another machine, and the epics, tasks, dependencies, and follow-ups are still intact.
• Git is the sync layer: commit .mycelium/ and the plan travels with the code. A teammate (or agent) who pulls the branch inherits the exact same task graph. No server, no account, no external service.
• One-shot and scriptable: every command is a single non-interactive invocation with --format json and --quiet for clean parsing. An agent drives it without a REPL or a session to manage.
• A drop-in agent contract: myc init writes an AGENTS.md describing the workflow, so the agent knows how to use the tool without you explaining it. Follow-ups let it jot "oh-by-the-way" findings mid-task without derailing.
• Blocking that means something: dependency links with cycle detection, so an agent can't close a task whose blockers are still open.

How it works in practice

A walkthrough of a real agent session in a repo.

1. Setup: the agent gets its instructions. Running myc init (once, per project) writes an AGENTS.md into the repo describing the whole workflow: the commands, the data model, and the rules (e.g. "every task belongs to an epic", "surface open follow-ups before wrapping up"). Agents like Claude Code read AGENTS.md automatically, so the tool is self-documenting and you never have to explain myc to the agent.

2. During work: the agent tracks as it goes. You ask the agent to build a feature. It decomposes the work into the plan, right in the repo:

myc epic create --title "Auth" --description "Login + sessions"
myc task create --title "Password login" --epic 1 --priority high
myc task create --title "Session middleware" --epic 1
myc task link blocks --task 2 3       # sessions depend on login
myc task update 2 --status in_progress

Mid-task it notices something unrelated and jots it down without losing focus, using the lightweight follow-ups scratchpad:

myc followup add "TODO: rotate the JWT secret, it's hardcoded in config.rs"

3. Wrap-up: nothing gets silently dropped. When the agent finishes, a Claude Code Stop hook (installed automatically by myc init) fires and checks for open follow-ups. If any exist, it feeds them back to the agent, which surfaces them to you instead of ending the turn as if everything were done:

Before we wrap, 1 open follow-up: "rotate the JWT secret, hardcoded in config.rs". Want me to handle it now, or leave it for later?

The hook self-gates: it stays silent outside mycelium projects and only fires when there's actually something open, so it never nags.

4. Next session: the state is still there. Days later, a fresh agent session (cleared context, or a teammate on another machine who pulled the branch) starts by reading the plan back, not by asking you what happened:

myc list                 # the Auth epic, its tasks, "Session middleware [blocked by T2]"
myc followup list -o     # the JWT-secret note is still waiting

It resumes exactly where the last session left off. The plan lived in .mycelium/mycelium.db and travelled with the code through git.

Features

• Agent-Optimized: One-shot CLI, --format json, --quiet, AGENTS.md contract
• State Persistence: Plan survives context resets; reconstructable from myc list
• Git-Trackable: SQLite storage designed for version control and branch sync
• Dependency Management: Task blocking with cycle detection
• Category-Prefixed IDs: Displays E3/T3/F3 so epic/task/follow-up IDs never get confused; input still accepts bare numbers
• Follow-ups: Lightweight scratch table for non-blocking items captured mid-work
• Smart List View: Tree visualization for dependencies, epic grouping for simple lists
• Assignees: Local assignees with GitHub username linking
• External References: Link tasks to GitHub issues/PRs and URLs
• Task Notes: Add comments and notes to tasks
• Task Cloning: Duplicate tasks with all metadata
• Batch Operations: Close, tag, or move multiple tasks at once
• Single Binary: Statically compiled, no dependencies
• Fast: Sub-100ms response time for typical operations
• Safe: Comprehensive error handling and validation

Installation

Every tagged release ships prebuilt myc binaries for macOS (arm64/x86_64), Linux (x86_64), and Windows (x86_64), so most methods below need no compilation. cargo install is the only one that always builds from source.

CLI (myc) via Homebrew (macOS/Linux, no compile)

brew install tcsenpai/tap/myc

CLI (myc) via cargo-binstall (no compile)

Downloads the prebuilt release binary instead of building it:

cargo binstall mycelium-manager

CLI (myc) via crates.io (builds from source)

cargo install mycelium-manager

Installs the myc binary to ~/.cargo/bin/. Requires Rust 1.75+. This is the only method that compiles from source — use cargo binstall above to skip it.

Updating

myc update

Updates the binary via cargo install --force, then resyncs this project's AGENTS.md and follow-up hook to the new version. If cargo isn't available it skips the binary step and just resyncs the artifacts (update the binary by hand, then rerun).

CLI (myc) from source

git clone https://github.com/tcsenpai/mycelium
cd mycelium
cargo build --release
# Binary will be at target/release/myc
sudo cp target/release/myc /usr/local/bin/

GUI (MycUI) pre-built download

Each tagged release ships desktop bundles on the GitHub Releases page: .dmg/.app (macOS), .deb/.AppImage (Linux), and .msi/.exe (Windows). Download the one for your platform and install it. Every release also attaches the myc CLI binaries, alongside the crates.io publish.

Note: macOS bundles are currently unsigned, so Gatekeeper may warn on first launch (right-click → Open to bypass).

GUI (MycUI) from source

MycUI is a Tauri desktop app built with React and TypeScript.

Prerequisites: Rust, Bun, and Tauri prerequisites for your platform.

cd mycui
bun install
bun run tauri:build

The built app will be in mycui/src-tauri/target/release/bundle/ with platform-specific installers (.deb, .AppImage, .dmg, .msi).

For development:

cd mycui
bun install
bun run tauri:dev

One-Line Install (Linux & macOS)

git clone https://github.com/tcsenpai/mycelium && cd mycelium && ./install.sh

The install script detects your platform and installs both myc (CLI) and MycUI (GUI). For the CLI it downloads the prebuilt release binary and only falls back to a source build if no matching asset exists. MycUI is always built from source (Tauri). On macOS, MycUI is installed as /Applications/MycUI.app; on Linux, both binaries go to /usr/local/bin/.

./install.sh --cli            # Install the CLI (prebuilt, falls back to source)
./install.sh --gui            # Install only MycUI (source build)
./install.sh --all            # Install both (default)
./install.sh --from-release   # Force the prebuilt CLI binary (no cargo needed)
./install.sh --build          # Force a source build of the CLI
INSTALL_DIR=~/.local/bin ./install.sh --cli   # Custom install path (CLI)
MYC_VERSION=0.4.2 ./install.sh --from-release  # Pin a specific prebuilt version

Versioning

The CLI (myc) and the desktop app (MycUI) are versioned independently and released on their own cadence. The CLI is published to crates.io as mycelium-manager; MycUI ships as a desktop bundle attached to GitHub Releases. There is no requirement to run matching version numbers.

They stay compatible because both are built on the same mycelium-core crate, so they share one schema, one set of migrations, and one data layer. You can run any released CLI alongside any released MycUI against the same project, use either tool on its own, or mix them freely. Neither depends on the other at runtime: the CLI is a standalone binary, and MycUI talks to the database directly through core rather than shelling out to myc.

Quick Start

# Initialize a new mycelium project
myc init

# Create an epic
myc epic create --title "Feature X" --description "Build feature X"

# Create tasks
myc task create --title "Design API" --description "Define the API surface and contracts" --epic 1 --priority high
myc task create --title "Implement backend" --description "Build the backend services and persistence layer" --epic 1 --priority critical --due 2025-06-01

# Set up dependencies (task 1 blocks task 2)
myc task link blocks --task 1 2

# View dependency tree
myc deps show 2

# Close tasks (blocked tasks prevent closing)
myc task close 1
myc task close 2

# Batch operations
myc task batch-op close 3 4 5              # Close multiple tasks
myc task batch-op tag urgent 1 2 3         # Tag multiple tasks
myc task batch-op move 2 6 7 8             # Move tasks to epic #2

# Task notes
myc task note 1 "Found edge case with auth"
myc task notes 1                           # View notes

# Clone a task
myc task clone 1 --title "API Design v2"

# View project summary
myc summary

Commands

Project

myc init                    # Initialize mycelium (also installs the follow-up hook)
myc init --no-hooks         # ...without installing the hook
myc update                  # Update the binary + resync AGENTS.md and the hook
myc summary                 # Show project overview
myc doctor                  # Check system health and configuration

IDs are category-prefixed. Commands display E3 (epic), T3 (task), F3 (follow-up). You can pass either the bare number (myc task show 3) or the prefixed form (myc task show T3, case-insensitive); a wrong-category prefix (myc task show E3) is rejected with a hint. JSON/CSV keep raw integer IDs.

Epics

myc epic create --title "..." [--description "..."]
myc epic list               # List all epics with task counts
myc epic show <id>          # Show epic details with tasks
myc epic update <id> [--title "..."] [--description "..."] [--status open|closed]
myc epic delete <id> [--force]

Tasks

myc task create --title "..." [options]
  --epic <id>               # Assign to epic
  --priority <low|medium|high|critical>
  --assignee <id>           # Assign to person
  --due <YYYY-MM-DD>        # Set due date

myc list [filters]          # List tasks and epics (shows tree view if dependencies exist)
  --epic <id>               # Filter by epic
  --status <open|closed>    # Filter by status (defaults to 'open')
  --priority <level>        # Filter by priority
  --assignee <id>           # Filter by assignee
  --blocked                 # Show only blocked tasks
  --overdue                 # Show only overdue tasks
  --all                     # Show all tasks including closed

myc task list [filters]     # Same as myc list, but task-specific

myc task show <id>          # Show task details
myc task update <id> [options]
myc task close <id> [--force]   # Close (blocked tasks need --force)
myc task reopen <id>
myc task delete <id> [--force]
myc task assign <task_id> <assignee_id|0>

Batch Operations

# Close multiple tasks at once
myc task batch-op close <id> [<id>...] [--force]

# Add a tag to multiple tasks
myc task batch-op tag <tag> <id> [<id>...]

# Move multiple tasks to an epic (use 0 for no epic)
myc task batch-op move <epic_id> <id> [<id>...]

Task Notes

myc task note <task_id> "Note content"    # Add a note to a task
myc task notes <task_id>                  # Show all notes for a task

Task Cloning

myc task clone <id> [--title "New Title"]  # Clone a task (copies description, priority, etc.)

Dependencies

myc task link blocks --task <blocker_id> <blocked_id>
myc deps show <task_id>     # Show dependency tree
myc deps unlink <task_id> <blocked_task_id>

Assignees

myc assignee create --name "..." [--email "..."] [--github "username"]
myc assignee list           # List with task counts
myc assignee show <id>
myc assignee delete <id> [--force]

External References

myc task link github-issue --task <id> "owner/repo#123"
myc task link github-pr --task <id> "owner/repo#456"
myc task link url --task <id> "https://..."
myc task unlink <ref_id>

Follow-ups

Lightweight scratch table for non-blocking items captured mid-work (bugs you noticed, questions, ideas). They are separate from tasks and carry no epic, priority, or dependencies. Body is required, title optional. Statuses: open, in_progress, done, wontfix.

myc followup add "body text" [--title "tag"]    # capture
myc fu add "short form alias works"

myc followup list                               # all items (default, alias -a)
myc followup list -o                            # only active (open + in_progress)
myc followup list -c                            # only closed (done + wontfix)
myc followup list --status done                 # exact status bucket
myc followup show <id>
myc followup next                               # lowest-ID active

myc followup start <id>                         # → in_progress
myc followup done <id> [--reason "..."]
myc followup wontfix <id> [--reason "..."]
myc followup reopen <id>

myc followup edit <id> --body "new body" [--title -|"new title"]
myc followup append <id> "more context"         # timestamped, additive
myc followup rm <id> [--force]
myc followup promote <id> [--epic N] [--priority high]   # convert to task
myc followup count                              # JSON: {open, in_progress, done, wontfix}
myc followup snooze [--turns N]                 # silence the Stop hook for N stops (default 5)
myc followup snooze --turns 0                   # clear an active snooze

After myc task close, mycelium prints a one-line reminder if any active follow-ups exist. Agents using mycelium MUST run myc followup list at the end of every work unit and surface open items to the user before wrapping (see AGENTS.md).

The follow-up Stop hook re-checks on every stop while open items exist. Once you've surfaced them and the user decides to leave them for later, myc followup snooze silences the hook for the next few stops (project-scoped, one stop consumed per snooze) instead of re-prompting each turn.

Hooks

myc hooks install            # install the follow-up Stop hook into .claude/ (local)
myc hooks install --global   # ...into ~/.claude/ instead
myc hooks uninstall          # remove it (add --global for ~/.claude)
myc hooks status             # show where it's installed

Reporting & Export

myc summary                 # Project overview
myc export json [--output file.json]
myc export csv [--output file.csv]

Global Options

--format <table|json>       # Output format (default: table)
--quiet                     # Suppress non-error output
--help                      # Show help
--version                   # Show version

Data Model

Epic
├── id, title, description
├── status (open/closed)
└── Tasks[]

Task
├── id, title, description (optional)
├── status (open/closed), priority (low/medium/high/critical)
├── epic_id (optional), assignee_id (optional)
├── due_date (optional)
├── dependencies (blocks/blocked_by)
└── external_refs (GitHub issues/PRs, URLs)

Git Integration

Mycelium stores data in .mycelium/mycelium.db using SQLite with WAL mode. This makes it git-trackable:

# Add to your repo
git add .mycelium/
git commit -m "Add mycelium project tracking"

# The .mycelium/.gitignore excludes WAL files

For AI Agents

Mycelium is optimized for agentic workflows.

The resume pattern. At the start of a task, the agent reads its own prior plan instead of relying on the context window:

myc list --format json        # what's the current task graph?
myc task list --blocked       # what's stuck?
myc followup list -o          # any open "oh-by-the-way" items from last time?

At the end, it records state that will still be there next session:

myc task update 4 --status closed
myc followup add "auth tests are flaky under load, investigate"

Across a context reset or a fresh session on the same branch, the plan is unchanged and the agent picks up exactly where it left off.

Everyday commands:

# Use --quiet to get just IDs
myc task create --title "New task" --quiet  # outputs: 42

# Include a description when useful
myc task create --title "New task" --description "Explain the work item"

# Use --format json for parsing
myc task list --format json

# Check blocked tasks
myc task list --blocked

# Export for analysis
myc export json

Claude Code follow-up hook

A Claude Code Stop hook enforces the end-of-task follow-up check, so it no longer relies on the agent remembering the AGENTS.md rule. The hook script is embedded in the myc binary, so it ships with cargo install — no extra files to fetch.

myc init installs it automatically into the project's .claude/ (commit that directory so the whole team gets the check). Manage it explicitly with:

myc hooks install            # project-local .claude/ (default)
myc hooks install --global   # ~/.claude/ instead
myc hooks uninstall          # remove (add --global for ~/.claude)
myc hooks status             # show where it's installed
myc init --no-hooks          # initialize without installing the hook

The hook self-gates: it stays silent outside mycelium projects (detected via the myc:agents-start marker in AGENTS.md) and only fires when open follow-ups exist. It also self-dedups on the session, so a global and a project-local copy can coexist without firing the check twice. Requires jq.

Prefer a shell installer for a global-only setup? hooks/install-hook.sh (and --uninstall) still installs into ~/.claude/ without the binary.

Configuration

No configuration needed! All data is stored in the project-local .mycelium/ directory.

Safety Features

• Atomic operations: Database transactions ensure data integrity
• Dependency validation: Circular dependencies are prevented
• Blocker checks: Tasks with open blockers cannot be closed (without --force)
• Confirmation prompts: Destructive operations require --force or user confirmation
• Clear errors: All errors include actionable guidance

Performance

• Sub-100ms response time for typical operations
• SQLite with proper indexing
• WAL mode for concurrent read/write
• Single binary, no runtime dependencies

Development

# Clone
git clone https://github.com/tcsenpai/mycelium
cd mycelium

# Build & test the CLI
cargo build --release
cargo test
cargo run -- init

# Run MycUI in dev mode
cd mycui
bun install
bun run tauri:dev

Architecture

• Rust - Type-safe, performant, single binary
• mycelium-core - Shared crate (db, models, migrations, errors) consumed by both the CLI and MycUI, so they use one schema and data layer
• SQLite - Embedded, git-trackable, ACID-compliant
• Clap - Command-line parsing with derive macros
• Rusqlite - SQLite bindings with bundled lib
• Tauri - Desktop GUI framework (MycUI)
• React + TypeScript - MycUI frontend
• Hand-written CSS - MycUI styling (no CSS framework)

License

MIT License - see LICENSE file.

Contributing

Contributions welcome! Feel free to open issues and pull requests.

Acknowledgments

Inspired by beads and the need for a robust, git-trackable task manager that works seamlessly with AI agents.