Pi AI 0.99 Adds MCP, Codemode, Tool Search

// 1h agoPRODUCT UPDATE

Pi AI 0.99 Adds MCP, Codemode, Tool Search

Pi 0.99 adds built-in MCP, model-written JavaScript workflows through Codemode, and on-demand tool search, alongside virtual and classifier model support. The release makes Pi a more capable, extensible terminal agent while preserving its open-source, customizable architecture.

// ANALYSIS

Pi 0.99 turns extensibility into a first-class agent runtime: models can discover tools selectively and compose multi-step workflows in code, though teams still need to consider sandboxing, permissions, and observability.

–Built-in MCP connects external servers over stdio or streamable HTTP, with configuration and authentication managed inside Pi.
–Codemode runs JavaScript in a QuickJS sandbox, enabling parallel tool calls and filtering intermediate results before they reach the model.
–Tool search defers undeclared tools until needed, reducing the context burden of large tool collections.
–Virtual and classifier models let extensions route requests dynamically, pushing Pi beyond a conventional coding CLI.
–Its open-source, replaceable architecture remains the differentiator: teams can adopt the bundled defaults or swap them through extensions.

// TAGS

pi-aicoding-agentai-codingagentmcptool-usecliopen-source

DISCOVERED

1h ago

2026-10-02

PUBLISHED

1h ago

2026-10-02

RELEVANCE

8/ 10

AUTHOR

DIY Smart Code

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

RESEARCH26m ago

RADAR Detects, Tames Reasoning Loops

A new paper introduces RADAR, a real-time method for detecting when large reasoning models drift from productive reflection into redundant or persistent generation loops. Attention realignment reduces these failures while largely preserving normal performance.

RESEARCH26m ago

CICM exposes stale binding in LLMs

UC Berkeley researchers introduce CICM, a benchmark showing that LLMs can retain updated facts yet answer with outdated values because attention selects stale mentions. The paper proposes training-free attention redirection that corrects most tested errors.

RESEARCH26m ago

Long-Transduction exposes long-horizon agent failures

NVIDIA researchers introduce Long-Transduction, a controlled diagnostic for testing whether models can repeatedly read, mutate, and output state-dependent results across long contexts. Across open-weight models, performance fell 62.8% as context expanded from 4K to 128K, with additional losses from input-format variation and task complexity.