PulSeed

Keep a goal moving until it is actually done.

PulSeed is an AI orchestrator for work that does not finish in one chat turn. Tell it the outcome you want, and it keeps observing, delegating, verifying, and looping until the goal is reached or the plan needs to change.

The primary entry point is pulseed. The normal flow is natural language, not a menu of subcommands.

Get Started

PulSeed supports Node.js 22 or 24.

Quick install (macOS / Linux):

curl -fsSL https://raw.githubusercontent.com/my-name-is-yu/PulSeed/main/scripts/install.sh | bash

Quick install (Windows / PowerShell):

irm https://raw.githubusercontent.com/my-name-is-yu/PulSeed/main/scripts/install.ps1 | iex

Then start PulSeed:

pulseed

For pinned-tag installs, fallback npm installs, and installer flags, see Getting Started.

Then describe the goal in natural language:

• Increase test coverage to 90%.
• Show me the current progress.
• Keep this goal moving in the background.

PulSeed will guide provider and adapter setup when needed.

Links

• Get Started
• Docs Index
• Use Cases
• Runtime
• Configuration
• Status
• Contributing
• Security

Why PulSeed?

• Goal-first orchestration for long-running work
• Honest goal negotiation when a target is not realistic as stated
• Bounded agent execution with verification around delegated work
• Verified playbook memory that turns successful runs into reusable planning hints
• Local persistent state under ~/.pulseed/
• Multiple runtime surfaces: CLI, chat, TUI, daemon, and cron
• Support for OpenAI, Anthropic, Ollama, and adapter-based execution paths

What It Does

• CoreLoop keeps a goal moving and decides whether to continue, refine, verify, or stop
• AgentLoop handles bounded tool-using work for task execution, chat, and selected runtime phases
• Centralized AgentLoop profiles keep task execution isolated by default, narrow chat permissions, and run /review through a dedicated read-only review posture
• Dream-backed playbooks can feed verified workflow hints into later task generation without auto-writing executable skills
• State, reports, schedules, and local memory live under ~/.pulseed/
• Software-level approval and verification gates protect delegated work

Common Surfaces

• pulseed for the primary interactive workflow
• pulseed tui for the terminal UI
• pulseed start and pulseed stop for daemon control
• pulseed schedule ... for schedule management
• Lower-level commands for scripting, diagnostics, and compatibility

Use Cases

PulSeed is a fit when the work is bigger than one prompt and needs follow-through:

• keep a codebase moving toward a maintenance goal
• track a business KPI and adjust strategy as results change
• coordinate recurring reports, reminders, and external actions
• observe external data sources and ask for confirmation when a human decision is required

More detailed examples live in Use Cases.

Docs and Community

Start with the public doc map:

• Getting Started
• Runtime
• Mechanism
• Configuration
• Architecture Map

For project participation:

• read Contributing before opening a pull request
• use Issues for bugs and feature proposals
• follow the Code of Conduct

Safety Boundary

PulSeed uses approval gates and verification around delegated work. Native agent_loop task execution can use isolated git worktrees, and supported CLI adapters can be wrapped with a Docker terminal backend. These reduce blast radius, but local backends and plugins still run with the user's privileges. See Security.

License

MIT