Pi 1.0

Pi 1.0 (Earendil) is a terminal-based AI representative harness launched by Earendil on October 1, 2026. It works as a very little, extensible platform through which users can run AI representatives that integrate thinking, tool invocation, and design changing. Its function is to assist company experts automate workflows, carry out code or scripts, procedure files, create images, and chain jobs– all through a combined user interface that enables personalization and combination with several LLM suppliers. Pi 1.0 targets those who desire agentic, programmable AI workflows with fine-grained control over tooling and habits.

Secret Features

• Representative Harness & & Codemode: Pi 1.0 provides a representative harness that supports codemode– a mode in which the AI design composes JavaScript code to conjure up tools from a sandbox environment instead of running in a simple prompt-reply loop. The codemode execution is leaner in variation 1.0, lowering timely size for particular jobs. It likewise supports image generation through the models.generateImages() API.

• Deferred Tool Loading & & Virtual Models: New in Pi 1.0 is assistance for delayed loading of tools (tools are just packed when required), and virtual design extensions which permit users to make up or cover several designs or habits into customized “virtual designs.”

• Cache Warming & & Mid-Conversation Changes: To enhance efficiency, Pi 1.0 includes cache warming for specific designs( for instance, those from Anthropic ), along with assistance for customizing system guidelines or tool sets mid-conversation– permitting representatives to adjust their policies or tool abilities gradually.

• Minimalism & & Extensibility: Pi is open-source under the MIT license. It is developed to be light-weight, flexible, and personalized. Designers can construct extensions, specify representatives with profiles, specify tools, handle service providers, set thinking levels, and incorporate lots of design backends. Pi likewise consists of a buddy speculative plan called “Pi Durable” for long-running agentic applications with long lasting state and multi-surface interaction.

Prices

Pi itself is open-source under an MIT license, implying the core representative harness is complimentary to utilize and can be extended or released commercially without paying simply for the MIT-licensed software application. Token use and expenses emerge from the design service providers you link: OpenAI, Claude, Radius, DeepSeek, and so on. Various designs have various prices per input token, output token, cache read, and cache compose. The Radius service provider’s “Balanced” design is priced at roughly 3 USD per million input tokens, 15 USD per million output tokens, 0.30 USD per million cache-read tokens, and absolutely no for cache composes. There are no set month-to-month membership tiers released for Pi 1.0 itself; rather, expenses are use based, based on which design supplier and design you utilize, the number of tokens you take in, and just how much tool or classifier invocation takes place. If you utilize Pi with regional designs, or your own hardware, service provider expenses might be very little or restricted to facilities.

Who is it for?

Pi 1.0 is well matched for designers, engineers, information groups, and technical decision-makers who wish to automate workflows or embed smart representatives into organization procedures. It is specifically suitable when:

• You require custom-made tooling or scripting– for instance, immediately producing reports, changing information, or carrying out code based upon file contents.

• You have multi-step agentic workflows needing numerous design service providers, selective tool use, or vibrant design changing (e.g. preparation in one design, execution in another).

• You choose open-source, self-hosted or extendable systems instead of closed, SaaS-only representatives, and desire control over what tools, service providers, or designs are utilized.

• Your work includes bursty use or project-based use where token-based billing (pay-as-you-go) is more versatile than repaired memberships.

Pi 1.0 may be less ideal for users who choose low-latency, turn-key AI assistants without setup, or those who are less comfy handling tool meanings, design qualifications, or extensions.

Last ideas

Pi 1.0 marks a considerable maturity turning point for Earendil’s representative harness innovation: steady, solidified, and developed with a user-first method to extensibility and efficiency. It strikes a balance in between power and very little intricacy, specifically for organizations that require programmable representatives instead of nontransparent assistants. For any company with designer capability and a requirement for versatile AI representatives, Pi 1.0 provides strong worth. The reward is greatest when you thoroughly select your design suppliers, enhance tool use, and handle token expenses, because prices is usage-based instead of membership based.

Go to the main site for more.

Maintain to date with our stories on LinkedInTwitterFacebook and Instagram