Pi 1.0 | Earendil


Today we’re proudly transport Pi 1.0: a hardened, minimal, extensible agent
harness which you could make your personal. Hundreds of hundreds of individuals across the
international stage use Pi each week. Many of you submit points and pull requests. Over the
course of many months, we now have used that suggestions to enhance, harden and evolve
Pi right into a steady piece of software program that individuals and companies can depend upon.

Pi is understood for being minimal. We care about holding that line. While agentic
tooling modifications each week, lots of the modifications don’t final. Pi doesn’t work
that method. We wait till one thing has confirmed itself, and solely then will we
contemplate adopting it; weighing its true performance towards its inherent
added complexity. We spoke extra about that course of and the way it pertains to
Codemode and MCP earlier this week here.

The Pi 1.0 we’re transport immediately is a results of that course of. Pi already ran
the most recent fashions from each main supplier, turned a day by day driver coding agent
for individuals across the international stage, and supplied a supermalleable substrate on which
to construct agentic purposes. With Pi 1.0 we’re including the next into Pi:

  • Codemode (native help for MCP, and non-LLM fashions like Jev and picture fashions)
  • Extension help for digital fashions
  • Deferred device loading
  • Cache warming for anthropic fashions
  • Mid-conversation system messages (transcript-aware immediate and power modifications)
  • A brand new TUI theme
  • Full-screen mode by default

We have considered many of those options for months. They had been thrown up
towards the wall and caught. The record of issues that fell off the wall is way
longer. When we use Pi with all these new options it seems like a serious step
ahead, but it surely additionally nonetheless feels easy, like Pi.

Alongside this work of regularly refining and evolving Pi, we additionally started to
perceive that there have been elements of Pi that didn’t match the form of what number of
individuals wished to make use of it. At Earendil, we need to convey Pi’s minimalism and the
method it permits you to wield AI exterior of the coding agent and outdoors of the
terminal. It must be reachable from completely different surfaces and help
longer-running conversations and duties. In quick, Pi should change into extra sturdy.
Rather than stray from our minimal roots and attempt to make Pi one thing it
wasn’t, we mixed that work into a brand new experimental bundle that we’re
transport immediately, Pi Durable. Pi Durable is a brand new substrate for constructing
long-running agentic purposes that lets the builders and the customers of these
purposes wield and steer the underlying intelligence with super
dexterity. It shares Pi’s key rules, minimalism and supermalleability, however
it extends these rules into new dimensions. Mario writes in additional element
about how we crafted Pi Durable here.

Earendil was shaped to craft software program and open protocols that strengthen human
company. Today we’re transport instruments to you all to assist just do that. We’re
keen to listen to suggestions from you and to harness the longer term out within the open
collectively.

Pi 1.0 is on the market immediately:

curl -fsSL https://pi.dev/install.sh | sh

or on Windows

powershell -c "irm https://pi.dev/install.ps1 | iex"

Pi Durable is on the market immediately as an experimental bundle:

npm set up @earendil-works/pi-durable @earendil-works/pi-ai @earendil-works/chord

Both are MIT licensed. Documentation is at pi.dev. Code is at
github.com/earendil-works/pi.

  1. Codemode: Pi writes a script that boils every week of commits all the way down to a brief abstract
  2. Pi writes itself an extension: a digital mannequin that plans with Claude Opus and implements with GPT, with Jev deciding when to change
  3. Reload, begin a brand new session, and choose the brand new router/auto mannequin
  4. Planning runs on Claude Opus
  5. Jev spots the change to implementation and fingers over to GPT 6 Luna
  6. ! runs the completed script
  7. /session breaks down price per mannequin and cache use



Source link