Tools

Agent workflow tools.

A public inventory of extensions for AI coding-agent work: context rollover, browser automation, provider integration, goal tracking, delegation, and model handoffs.

Why these tools exist

Coding agents are powerful. Real workflows still need infrastructure.

Goals and state disappear across long sessions, browser and desktop context sit outside the coding loop, providers are split across tools, tool activity needs inspection, hard decisions need async second-model review, and artifacts should be recoverable instead of scattered.

Current working suite

  • pi-posthorse moves long sessions to fresh context without lossy summaries.
  • pi-cursor-sdk brings Cursor-backed models into Pi’s workflow surface.
  • pi-agent-browser-native gives Pi agents browser and desktop-app automation.
  • pi-codex-goal keeps long-running agent sessions anchored to a goal.
  • pi-oracle sends context-rich web-model jobs without blocking the main session.

5 focused tools with current public GitHub and npm package signals. Downloads are package telemetry, not verified active users.

No-summary context rollover for Pi, with durable notes and history recovery.

Workflow gap
Long agent sessions hit the context limit, and summary-based compaction replaces the work so far with a lossy recap.
What it adds
Moves the agent to a fresh context window with a handoff it writes itself, calls no summary model, keeps the full transcript searchable, and adds notes that survive resets.
Public package signal
256 GitHub stars, 7 forks; npm 0.8.0; 1,415 npm downloads last month.
  • Pi
  • context windows
  • session history
  • durable notes

Cursor SDK provider extension for Pi.

Workflow gap
Cursor-backed agents and models should be usable from the same controlled workflow surface as other coding-agent providers.
What it adds
Brings Cursor models into Pi’s provider/model flow, with model-picker integration, local Pi tool bridging through MCP, and Pi-native tool activity replay.
Public package signal
339 GitHub stars, 75 forks; npm 0.5.0; 9,948 npm downloads last month.
  • Pi
  • Cursor
  • provider integration
  • MCP
  • model workflow

Native browser and desktop-app automation for Pi agents.

Workflow gap
Real agent work often needs browser context, screenshots, clicks, downloads, QA checks, and app inspection without leaving the coding session.
What it adds
Exposes agent-browser as a native Pi tool with browser automation, compact snapshots, screenshots, downloads, QA presets, source lookup, and Electron lifecycle support.
Public package signal
232 GitHub stars, 38 forks; npm 0.9.3; 25,067 npm downloads last month.
  • Pi
  • browser automation
  • Electron
  • QA
  • tooling

Codex-style goal tracking and continuation for long-running Pi sessions.

Workflow gap
Long-running coding-agent sessions lose their objective, drift across compaction, and become hard to resume cleanly.
What it adds
Adds a concrete goal object, token/progress visibility, and continuation context so agent work can stay anchored across long sessions.
Public package signal
195 GitHub stars, 29 forks; npm 0.6.0; 4,851 npm downloads last month.
  • Pi
  • Codex
  • goal tracking
  • continuation
  • session state

ChatGPT/Grok web-oracle extension for Pi with isolated browser auth and async jobs.

Workflow gap
Hard coding decisions sometimes need another model with full project context without blocking the main agent session.
What it adds
Packages project context, submits web-model jobs, tracks async status, archives artifacts, and reads results back into the workflow.
Public package signal
39 GitHub stars, 11 forks; npm 0.7.21; 637 npm downloads last month.
  • Pi
  • ChatGPT
  • Grok
  • async jobs
  • model handoff

Activity

Current public package signal

GitHub metrics refresh from public APIs; npm downloads last verified October 5, 2026 MDT.

  • pi-posthorse: 256 GitHub stars, 7 forks; npm 0.8.0; 1,415 npm downloads last month.
  • pi-cursor-sdk: 339 GitHub stars, 75 forks; npm 0.5.0; 9,948 npm downloads last month.
  • pi-agent-browser-native: 232 GitHub stars, 38 forks; npm 0.9.3; 25,067 npm downloads last month.
  • pi-codex-goal: 195 GitHub stars, 29 forks; npm 0.6.0; 4,851 npm downloads last month.
  • pi-oracle: 39 GitHub stars, 11 forks; npm 0.7.21; 637 npm downloads last month.

Downloads are package telemetry, not verified users.

Small focused extensions

Narrow utilities for session history and draft handling.

A pi extension for stashing draft messages and restoring them later.

Workflow gap
Draft prompts and partial thoughts are easy to lose while steering a live agent session.
What it adds
Adds a small stash/restore workflow for user drafts inside pi.
Public package signal
Published npm package with active public repo.

pi extension that lets you re-edit or delete an earlier user message in the current session branch.

Workflow gap
A bad earlier prompt can poison a session even when the rest of the branch is useful.
What it adds
Adds guarded in-place session message editing and deletion for pi branches.
Public package signal
Published npm package with active public repo.

Workflow infrastructure

Delegation, tool discovery, and native app automation for the rest of the agent workflow.

Async subagent delegation for Pi.

Workflow gap
Large tasks need specialist reviewers, parallel workstreams, and background work without flooding the main session's context.
What it adds
Delegates to specialist agents, runs parallel reviews, and keeps work going in the background, with truncated output and saved artifacts.

Token-efficient MCP adapter for Pi.

Workflow gap
Loading every MCP tool definition upfront spends context before the agent does any work.
What it adds
Connects Pi to external tools and services, discovering MCP tools as needed instead of loading every definition upfront.

pi / macos

macuse

Native macOS Computer Use for Pi and MCP clients. Experimental, unsupported integration.

Workflow gap
Some agent work happens in native macOS apps that browser automation cannot reach.
What it adds
Adds persistent JavaScript, guarded app actions, and selected-text insertion by reusing the proprietary Computer Use runtime installed with ChatGPT.

Where this is going

Pi is where this is getting traction now. The work is broader.

The work centers on practical infrastructure around coding agents: control, continuity, inspection, provider routing, browser access, recoverable artifacts, and second-model workflows. Pi is the visible distribution surface right now; the larger direction spans Cursor, Codex-style sessions, browsers, and multi-agent operations.