★ Overview
Cole Medin dedicates an entire livestream to exploring Pi, the minimalist coding agent built on the philosophy of "there are many agent harnesses, but this one is yours." Frustrated by Anthropic's increasingly harsh rate limits, Cole sets up Pi with a Kimi K2.6 subscription ($40/month), installs extensions from the marketplace, loads Archon skills, and live-builds a custom Archon Dispatch extension that turns Pi into a control panel for background Archon workflows. The stream demonstrates Pi's extensibility, the power of harness engineering, and the honest trade-offs of using cheaper models vs frontier models like Opus.
1 Why Pi? The Rate Limit Problem
Cole opens with the core motivation: rate limits are getting worse. He hit his Anthropic weekly limit in just 3 days the previous week — something that had never happened before. The industry trend is clear: frontier model subscriptions are getting more restrictive while models get more expensive.
- Pi's pitch — a minimal coding agent you build on top of, rather than a bloated tool you retrofit to your workflow
- "We adapt Pi to our workflows, not the other way around" — the core philosophy aligns with Cole's harness engineering ethos
- Model flexibility — unlike Claude Code or Codex, Pi supports any provider out of the box: Kimi, MiniMax, Qwen through OpenRouter, Codex, even GitHub Copilot
- Speed — being minimal makes Pi noticeably faster than bloated coding agents for simple requests
2 Pi's Philosophy & The Archon Connection
Cole explains how Pi fits perfectly with his existing work on Archon, his open-source tool for building custom agent harnesses. Both tools share the same philosophy: give developers control over their workflows instead of forcing them into opinionated systems.
- Archon integration — Archon already supports Pi as an AI assistant client alongside Claude and Codex
- Harness engineering — the harness (workflow, rules, extensions) matters more than the model for getting good results
- Dark Factory potential — Cole considers migrating his self-evolving codebase project to Pi because it can be molded to his specific workflow for triaging issues, handling PRs, and weekly regression testing
- Goal — build Pi to the point where he can reliably get Opus-level results with Kimi K2.6 through harness engineering alone
3 Model Providers — A Ton of Choices
Pi supports a massive range of model providers right out of the box — no workarounds needed like with Claude Code:
- Kimi for Coding — subscription-based ($40/month) with API key access, 5-hour and weekly rate limits
- MiniMax M2.7 — similar subscription model, another cheaper alternative
- Qwen 3.6 — available through OpenRouter, with impressive 27B and 35B parameter options
- Codex — Cole's primary provider up to this point
- GitHub Copilot — enterprise-oriented but supported
- Anthropic — technically possible but goes to extra usage (pay-per-token) not the flat rate subscription
4 Setting Up Kimi K2.6 — The API Key Incident
Cole walks through configuring Pi to use Kimi K2.6 instead of Codex. Pi demonstrates its meta-reasoning ability — it reads its own documentation to understand how to configure a new provider and creates the models.json file.
The Hilarious API Key Leak
After pasting his Kimi API key off-camera, Pi promptly displayed the entire key in plain text on the livestream when it read the file. Cole had to scramble to delete the key and create a new one on-camera.
After the key reset, Kimi was working perfectly. Cole confirmed via the Kimi console that requests were hitting his subscription, and the 262,000-token context window was available.
5 The Extensions Marketplace
Pi's extension marketplace is a core differentiator. Cole installs and tests several extensions live:
Pi Web Access
Installed with pi install npm:pi-web-access. Provides web search, URL fetching, GitHub repo cloning, PDF extraction, and YouTube video understanding. Without it, Pi creates janky Python scripts to make HTTP requests.
Pi Agent Extension (Bundle)
A multi-feature extension pack that includes:
- Status Line — fixed two-line footer showing model, thinking, context/token usage, work tree, and skill indicators
- Permission Gate — configurable danger command gate (like Claude Code hooks, but as an extension)
- Slow Mode — review gate for write/edit tool calls; approve/reject diffs before they hit disk
- Notify — desktop notification when the agent finishes a task
- Questionnaire — LLM-callable multi-question tool with tabbed UI (like Claude Code's ask_user)
- Stash — save/restore editor text
- Env autoload — loads .envrc environment variables into every bash call
6 The Skills System — Loading Claude Code Skills
Cole discovers how Pi's skills system works and how to make it compatible with existing Claude Code skills:
- Native skill discovery — Pi loads skills from
.agents/skills/and.pi/skills/directories - No native .claude support — Kimi hallucinated that Pi supports
.claude/skills/directly. It doesn't. - Workaround 1 — copy the
.claudefolder as.agents— works immediately - Workaround 2 (better) — add to
settings.json:{"skills": [".claude/skills"]}— Pi then loads Claude Code skills without renaming - Pi updates its own config — when asked, Pi read its own documentation, understood the settings.json schema, and made the config change itself
After configuration, Cole successfully ran an Archon workflow through Pi using the loaded Archon skill — confirming the full skill → CLI → workflow pipeline works.
7 Building Archon Dispatch — A Custom Extension
The centerpiece of the stream: Cole has Pi build a custom extension called Archon Dispatch that turns Pi into a control panel for Archon background jobs. He prepared a prompt using his "second brain" (separate Claude conversation for planning).
Extension Architecture
The Archon Dispatch extension includes five components:
- Archon Dispatch tool — a callable tool that spawns Archon workflows using Node.js
spawnwith shell mode - Confirmation gate — intercepts every tool call; pops a confirmation dialogue showing workflow name, branch, and message preview
- Live status line — maintains in-memory registry of running workflows; displays "N workflows running" with latest log line
- Progress tailing — tracks file position to never re-read old log data
- Completion loop — on process exit: updates status, fires desktop notification, injects completion message into conversation without forcing an LLM response (
delivermode vssendUserMessage)
Pi built the entire extension using Kimi K2.6 — less than 100,000 tokens, only 4% of the 5-hour Kimi rate limit and 1% of the weekly limit. The $40/month subscription proved its value.
8 Testing & Debugging — The Iterative Reality
Testing the Archon Dispatch extension reveals the real-world challenges of AI-built code:
Bug 1: Work Tree Flag
The extension always passed --branch to workflows, but Archon Assist has work_tree: false, causing failures. Fix: the tool now supports three modes — no work tree, branch, or neither.
Bug 2: Empty Log Files
Workflows ran successfully but log files were empty. Root cause: Node.js pipe buffering on Windows with detach: true lost the output. Fix: OS-level file descriptor redirection instead of piping through Node.js.
Bug 3: The "Lazy Kimi" Moment
Even after fixing the logging, Kimi decided to just tell Cole the log file path instead of reading and displaying its contents — despite being explicitly instructed to inject the full output into the Pi session.
Cole ultimately declared it a "partial success" — the extension dispatches workflows correctly, shows status, fires notifications, and injects completion messages. The log-reading issue would need Opus to fix properly.
9 Model Comparison — Cheaper Models vs Frontier
The livestream provides honest, real-time benchmarking of cheaper models:
- Kimi K2.6 ($40/month) — "Good and cheap, but not as good as I'd want. Always slightly disappointed." Gets stuck in reasoning loops, sometimes hangs entirely. Used only 8% of 5-hour limit and 2% of weekly limit for the entire stream.
- MiniMax M2.7 — similar performance tier, also hangs occasionally. Cole had good results in his Dark Factory.
- Qwen 3.6 (35B/3B active) — impressive benchmarks for its size, competitive with much larger models. Could run on a single 3090 GPU quantized.
- DeepSeek V4 Flash — 284B total, 13B active parameters. Very cheap. The "Pro Max" variant exists (Cole laughs at the naming).
- JCode — a new coding agent harness focused on speed/performance. Cole isn't impressed by speed alone: "Coding agents, you care the most about output quality over everything."
10 The RPIV Advisor — Dual-Model Architecture
Cole discovers a Pi extension called RPIV Advisor that he's excited to try in a future stream:
- Concept — a "second opinion" model that the working model can request reviews from before acting
- Architecture — working model (Kimi, Qwen, MiniMax) handles implementation; reviewer model (Opus) validates decisions and catches errors
- Why it matters — this is the kind of customization you can't do in Claude Code. It enables a cost-effective workflow where 90% of tokens are cheap but critical decisions get frontier-model review.
11 Final Verdict & The Road Ahead
Cole wraps up with honest assessments and future plans:
- Pi is genuine — the extensibility and model flexibility are real, not hype. Config changes, extension installs, and skill loading all work smoothly.
- Cheaper models are "pretty good" but disappointing — Kimi built most of the extension correctly but couldn't close the final bug. The reasoning loops and occasional laziness are real pain points.
- The sweet spot is combining models — use Opus for planning and review, cheaper models for implementation. Pi's architecture makes this natural.
- Next stream plans — fix the Archon Dispatch extension with Opus, explore the RPIV Advisor extension, potentially build an Archon workflow for systematically creating Pi extensions.
- Pi stands for "Prime Intellect" — a fun trivia fact revealed by a chat viewer.
🎯 Key Takeaways
🔑 Key Takeaways
- Pi is a foundation, not a finished product — it's intentionally minimal so you can build the coding agent that fits your exact workflow, unlike Claude Code or Codex which are opinionated and hard to customize
- Extensions are the killer feature — one-command installs from the marketplace add capabilities like web search, permission gates, questionnaires, and status indicators. Building custom extensions is also straightforward.
- Any model, any provider — Pi natively supports Kimi, MiniMax, Qwen, DeepSeek, Codex, Copilot, and more. No workarounds needed for alternative providers.
- Kimi K2.6 at $40/month is viable but imperfect — built an entire Archon Dispatch extension using only 4% of the 5-hour rate limit. However, reasoning loops, hallucinations, and "lazy" outputs remain real issues.
- Skills cross-compatibility — Pi can load Claude Code skills by adding the path to settings.json, no renaming required
- Meta-reasoning works — Pi successfully reads its own documentation, updates its own config, and builds extensions for itself
- Harness engineering > model size — the right workflow, rules, and extensions can make a weaker model produce Opus-quality results, at a fraction of the cost
- Dual-model architecture is the future — use cheap models for 90% of work, call Opus for critical reviews via the RPIV Advisor extension pattern
- Watch your API keys on stream — Pi (and any coding agent) will happily display sensitive data when reading files. Consider PII-redacting extensions.
- Anthropic rate limits are getting worse — the weekly limit was quietly reduced even as Anthropic announced "better" 5-hour limits. Pi + alternative models is a practical escape route.
- "The skill is the knowledge, the extension is the hands" — skills tell the agent what to do; extensions give it new capabilities for how to do it. Both are needed for a complete custom workflow.
🔗 Resources & Links
- Pi Coding Agent — official website with documentation and package catalog
- Pi GitHub Repository — open-source minimal coding agent (MIT license)