Recently, several noteworthy extensions have emerged in the Pi ecosystem. pi-discuss-mode lets you chat with your agent without worrying about accidental actions; Pi Exa adds web search capabilities to Pi; pi-workflow standardizes multi-agent workflows with a JSON spec. Here’s a collection of them.

Pi Extension Mini-Collection

pi-discuss-mode — Discuss safely with your Agent

Puts Pi into a “discussion only” mode. The agent can read code, answer questions, and make suggestions, but is restricted from any actions that modify files or run dangerous commands. Lightweight compared to “plan mode” — plan mode forces plan generation, while discuss mode lets you safely ask “how does this codebase work?”

/discuss # enable /discuss # disable

GitHub: [Still searching]

Pi Exa — Web search and deep research

Connects Pi to the Exa search engine. The agent can search the web, read pages, and conduct comprehensive research.

pi exa search “最新 JEPA 架构进展”

Show HN: https://github.com/junnjiee/pi-exa

pi-workflow — JSON workflow orchestration

Decomposes the sub-agent pattern into several stage types (split/orchestrate/build/review/combine), using a JSON spec to describe multi-agent collaboration. Comes with a skill to help write workflows.

{
  "stages": [
    {"type": "plan", "agent": "planner"},
    {"type": "orchestrate", "split": 3},
    {"type": "build", "parallel": true},
    {"type": "review", "fix": true},
    {"type": "combine"}
  ]
}

GitHub: https://github.com/AgwaB/pi-workflow…

Headroom × Oh My Pi — Model usage visualization

The headroom wrap command from Headroom now supports omp (Oh My Pi). You can see token usage per model, per transform, and per tool at the local proxy layer.

GitHub: headroomlabs-ai/headroom feat #1149

Community Voices

AI Jason (5.8K views): “My favorite part is parallel sessions — it lets 10-15 different agent sessions run safely and concurrently. Breaking down into 11 different agents with orchestrator built around task entity.”

#Pi #Extensions #AgentTools

junnjiee/pi-exa

Source: https://github.com/junnjiee/pi-exa

Pi Exa

Web search, fetching, and deep research for your Pi Agent. Powered by Exa (https://exa.ai).

No API key needed for web search and fetch. Works out of the box, install the extension and your agent can start using Exa search and fetch.

https://github.com/user-attachments/assets/8bc51126-f212-4ad7-813b-705197a972e8

Why Exa for your Agent

• Free - 1000 requests/month for web search and fetch, via the Exa MCP server (https://github.com/exa-labs/exa-mcp-server)
• Semantic search - understands intent, not just keyword searches
• LLM-friendly results - Exa returns clean snippets instead of dumping entire sites that nuke your context window (through Exa Highlights (https://exa.ai/blog/highlights-for-agents))
• Get what you want - Your agent can make sense of Exa’s search options and filters to get you narrowed results

Install

pi install npm:pi-exa

Optional: Exa API key for the deep search tool. To register your key:

pi /exa-login

Features

MCP

• Use web search and content fetching with no API key required
• Lazy-loaded Exa MCP server that only connects on first tool call
• Cached MCP tool schemas, Pi registers MCP tools during startup without connecting to the MCP server
• Toggle paid Exa MCP usage with /exa-mcp-use-api-key on|off

Advanced Web Search

• Let your agent use every Exa search filter
• Disabled by default to save context (~1200 tokens)
• Enable or disable with /exa-advanced-search on|off

Deep Search

• Give your agent access to Deep Search via the Exa API, with deep-lite, deep, and deep-reasoning modes
• Requires an Exa API key via /exa-login or EXA_API_KEY env var
• Enable or disable with /exa-deep-search on|off

Docs

Commands

Tools

Enabled while pi-exa is enabled:

Toggle to enable/disable while pi-exa is enabled. Preferences persist across sessions:

Bugs and suggestions

Open an issue!