AI agent toolkit: unified LLM API, agent loop, TUI, coding agent CLI
69.4k
Stars
8.5k
Forks
60
Open issues
100+
Contributors
AI Analysis
Pi is an AI agent toolkit providing a unified multi-provider LLM API (OpenAI, Anthropic, Google), an agent runtime with tool calling and state management, an interactive coding agent CLI, and a terminal UI library. It serves developers building or using LLM-powered coding agents and automation workflows directly from the terminal. It is not intended for end users unfamiliar with CLI tooling, nor for those seeking a sandboxed-by-default or GUI-driven AI assistant.
Inferred from signals mentioned in the README (tests, CI, type safety) — not a review of the actual code.
AI's overall editorial judgment — not an average of the bars above, can weigh other factors too.
Pi: A modular TypeScript agent harness with unified LLM API and coding-agent CLI, now one of the most-starred agent toolkits on GitHub
Pi is a monorepo agent toolkit built in TypeScript, offering a unified multi-provider LLM abstraction (OpenAI, Anthropic, Google, etc.), an agent runtime with tool calling and state management, a terminal UI library, and an interactive coding agent CLI. It targets developers who want to build or run AI coding agents without vendor lock-in. With 64K stars and active daily pushes as of June 2026, it has accumulated significant community attention very rapidly since its August 2025 creation. The project appears to be maintained by the earendil-works org with strong supply-chain discipline and a growing contributor ecosystem.
Created in August 2025, Pi is a young project that reached 64K stars in under a year. It appears to have grown out of earendil-works / badlogicgames' own agent development workflows, later open-sourced with a dedicated website and documentation.
Gaining nearly 1,900 stars in 7 days and 64K total in under 11 months suggests viral growth driven by the broader AI agent tooling wave of 2025-2026, a well-timed open-source release of a practical CLI coding agent, and active promotion via X/Hugging Face session sharing by the lead maintainer. The fork count (7,811) indicates meaningful developer engagement beyond passive starring.
The lead maintainer publicly shares real coding-agent sessions on Hugging Face, providing some evidence of genuine usage. A dedicated domain (pi.dev) and npm packages under @earendil-works are live. Fork count of 7,811 suggests real developer experimentation. However, large-scale enterprise production deployments are not documented in the README, and independent third-party case studies are not visible from the available metadata.
Appears to be a TypeScript monorepo (npm workspaces) with at least four packages: pi-ai (LLM abstraction), pi-agent-core (agent runtime), pi-coding-agent (CLI), and pi-tui (terminal UI). Likely uses a layered dependency structure where the coding agent depends on agent-core, which depends on pi-ai. The TUI library appears to be a standalone utility. Based on README, the design intentionally omits built-in sandboxing, delegating isolation to containers or VMs.
Partially documented: a test.sh script exists that skips LLM-dependent tests without API keys, and a pi-test.sh for integration runs. CI runs npm audit. The extent of unit vs integration test coverage is not fully documented in the README.
Last push was 2026-06-19, one day before the evaluation date — actively maintained. The README mentions daily maintainer review of auto-closed issues, RFC-based planning, supply-chain hardening practices (pinned deps, shrinkwrap, CI audits), and scheduled GitHub workflows. These are strong signals of an organized, sustainable maintenance posture for a project of this age.
ADOPT IF: you want a TypeScript-native, multi-provider LLM abstraction with a working coding agent CLI and are comfortable with the no-built-in-sandbox model (containerizing yourself). AVOID IF: you need enterprise-grade permission systems, Python-native tooling, or guaranteed long-term stability from an established vendor — Pi is young and the API surface may still shift. MONITOR IF: you're evaluating agent runtimes for a team product and want to see whether the ecosystem (plugins, integrations, RFCs) matures over the next two quarters before committing.
Independent dimensions
Mainstream potential
7/10
Technical importance
7/10
Adoption evidence
5/10
- No built-in permission or sandbox system — running untrusted agent-generated code requires manual containerization, raising the barrier for less experienced users.
- Extremely rapid star growth in under a year can attract contributors and forks that fragment the project or create ecosystem fragmentation (the oh-my-pi fork already has 13.5K stars).
- API surface likely still evolving given the project's age (< 11 months); downstream breakage risk is non-trivial for teams building on top of it.
- Contributor friction by design (auto-close of new issues/PRs) may slow community contributions and reduce bus-factor resilience if core maintainers deprioritize the project.
- No documented large-scale production deployments — adoption may be skewed toward experimentation and individual developers rather than verified production workloads.
Pi is likely to consolidate as a leading open-source TypeScript coding-agent toolkit over the next 12 months, provided the RFC roadmap delivers on extensibility and a sandboxing story emerges. Risk of fragmentation into competing forks is real.
Newsletter
Get analyses like this every Monday
Free weekly digest of the most interesting open-source discoveries.
Languages
Information
- Language
- TypeScript
- License
- MIT
- Last updated
- 31 min ago
- Created
- 11mo ago
- Analyzed with
- anthropic/claude-sonnet-4-6
Stars over time
Contributors over time
Top 100 contributors only — repos with more will plateau at 100.
Open issues
npm uninstall fails with ERESOLVE: peer dependency conflicts
Preserve unhandled Bedrock ConverseStream stop reasons in error messages
Allow disabling request compression for openai
Compaction summary requests omit the session ID, breaking compaction on some OpenAI-Codex models
Regression: httpIdleTimeoutMs no longer respected for self-hosted OpenAI-compatible provider (v0.80.6, worked fine in v0.80.3)
Top contributors
Similar repos
| Repository | Stars | Week Δ | Language | Score | Updated |
|---|---|---|---|---|---|
|
earendil-works/pi
this
|
69.4k | +1.9k | TypeScript | 8/10 | 31 min ago |
 can1357/oh-my-pi
|
17.1k | — | TypeScript | 8/10 | -50 min ago |
 wrtnlabs/agentica
|
1k | — | TypeScript | 7/10 | 2mo ago |
 pingdotgg/t3code
|
13.5k | — | TypeScript | 6/10 | 8h ago |
 1jehuang/jcode
|
8.3k | — | Rust | 8/10 | 8h ago |
 itayinbarr/little-coder
|
1.7k | — | TypeScript | 8/10 | 4d ago |
A fork/derivative with 13.5K stars, suggesting the ecosystem is fragmenting into variants — may indicate the core Pi design is being extended in different directions by the community.
A TypeScript agent framework with 9.7K stars, more framework-oriented vs Pi's CLI-first approach. Pi has substantially more adoption signals but VoltAgent may appeal to teams building server-side agents rather than developer tooling.
12.8K stars, TypeScript, appears to be a coding agent CLI in the same space. Likely more opinionated around a specific stack. Pi's multi-provider abstraction gives it broader applicability.
35.9K stars but is a Jupyter Notebook collection — a learning/reference resource rather than a competing runtime. Different audience entirely.
237 stars, TypeScript — a much smaller project in the same conceptual space. Not a meaningful competitive threat but signals the niche is attracting multiple independent efforts.
