Pi 1.0 Coding Agent: MCP via Codemode and Pi Durable Ships

Pi 1.0 landed on October 1, 2026 and immediately hit the top of Hacker News with 1,336 points. Earendil — the Vienna startup founded by Flask creator Armin Ronacher and backed by Accel — shipped two things at once: a polished stable milestone for its minimal Pi 1.0 coding agent, and Pi Durable, an experimental framework that checkpoints every agent move so a crashed container doesn’t cost you hours of work. Two features that have been on developers’ wishlists since multi-step agents became the default.

Pi 1.0 Coding Agent Finally Adds MCP — and the Implementation Matters

Pi previously rejected the Model Context Protocol on principle. Too much complexity, too many tokens. Earendil’s reversal at 1.0 comes with a clear rationale: “The best way to positively influence something is to embrace it.” What changed isn’t the philosophy — it’s the mechanism.

Codemode is how Pi adds MCP without eating its own advantages. Instead of loading all MCP tools permanently into context, Pi writes scripts to execute MCP tool calls on demand. Tools don’t sit idle in the system prompt; they enter only when the agent needs them. The result: MCP compatibility without the token sprawl Pi originally objected to. Codemode also extends to image models and decision models like Jev — it’s a scripting layer, not just an MCP wrapper.

The enterprise case for MCP ultimately won the argument. With server-side MCP, agents never see API keys directly — no credential leak in a compromised prompt. According to The Register’s coverage of the release, the HN discussion was split: power users who live in bash thought MCP was unnecessary overhead; enterprise developers thought credential isolation alone justified it. Both are right, depending on the deployment context.

Related: Claude Code Mods: Rewrite the Agent From the Inside Out

Pi Durable Solves the Problem Every Long-Running Agent Has

Every other CLI coding agent — Claude Code, OpenCode, Codex — terminates ungracefully. Run an agent for 20 minutes, hit a container restart, start over. That’s fine for a quick refactor. It’s not fine for an agent midway through a lengthy test suite, a large compilation, or a chain of sequential API calls. Pi Durable is the direct answer to that problem.

The architecture checkpoints everything: every model request, tool call, and compaction step is a durable task. On restart, interrupted model requests resend, partial responses get marked as aborted, and tool calls either rerun (if flagged “safe”) or report interruption to the model. Nothing disappears. Storage is pluggable — SQLite for local runs, JSONL for simpler deployments, with a small adapter that runs on Bun or inside a Cloudflare Durable Object.

const harness = await Harness.open(
    await openNodeSqliteStorage("./agent.sqlite"),
    { models, registry, env },
    context,
);
const root = await harness.root(context, {
    agent: { model: { provider: "openai", modelId: "gpt-6.1-sol" } }
});

Subagents are first-class in Pi Durable: each spawned conversation has its own model, tools, and cost tracking — and survives crashes independently. Multiple clients can watch and steer the same conversation via streaming deltas. Forking lets conversations branch at specific transcript points without duplicating history. Pi Durable is still experimental (~15,000 lines of source, API subject to change), so don’t ship production systems on it yet — but the architecture is mature enough to build against today.

The 1.0 Numbers: Cost, Tokens, and What Else Shipped

Pi’s benchmark position holds. According to Tensorlake’s 2026 coding agent comparison, Pi scores a 66.7% pass rate on a 30-task evaluation at $0.028 per task — against Claude Code’s $0.195. That’s 7x cheaper per task. For teams running thousands of agent jobs per month, that gap compounds into real money.

The 1.0 release also cuts Codemode prompt tokens by 40% — from 5,300 down to 3,300 tokens on GPT-5.6. Pi was already the leanest harness (~1,000 tokens system overhead vs Claude Code’s ~33,000), so cutting Codemode by 40% compounds the advantage further. Cache warming for Anthropic models, virtual model extensions for multi-model routing, and mid-conversation system message insertion complete the release. The full changelog is in the Pi 1.0 release notes.

However, the tradeoffs haven’t changed. Pi has no built-in sandboxing — run it in Docker or a VM when executing untrusted code. The TypeScript extension model requires setup work that more opinionated harnesses handle automatically. If your team needs guardrails without configuration overhead, Claude Code remains the safer default. Pi is for developers who know what they’re building and want the harness to stay out of the way.

Key Takeaways

• Pi 1.0 adds MCP via Codemode — tools load on demand rather than permanently in context, preserving the lean footprint the agent was built on
• Pi Durable introduces checkpoint-based crash recovery: agents survive container restarts, support multiplayer steering, and run on SQLite or inside Cloudflare Durable Objects
• At $0.028 per task versus Claude Code’s $0.195, the cost gap is 7x — significant for any pipeline running agent tasks at volume
• Pi Durable is experimental; build exploratory agents against it now, production deployments once the API stabilizes
• Multi-provider support (20+ providers, 300+ models) with mid-conversation switching remains Pi’s strongest differentiator against locked-in alternatives