Skip to content

Project

pi

Pi is an AI coding agent from Earendil Works that runs MCP tool calls in a sandbox, keeping their output out of the model's context.

Known aliases

  • @earendil-works/pi-coding-agent
  • Pi 1.0
  • Pi agent
  • Pi coding agent
  • Pi coding assistant
  • pi.dev

Current stories

buildOne report1 publisher

Pi 1.0 brings MCP into its core by batching tool calls in a JavaScript sandbox

Earendil's Pi 1.0, an agent harness whose website once advertised "No MCP", now builds MCP into its core. The team ties the reversal to Codemode, a JavaScript interpreter beside the harness loop that lets the model batch MCP calls into one script.

Publishers:dev.to

Reality

Evidence40
Adoption55
Hype gap+10
Incentives45
Confidence45
buildOne report1 publisher

Kubernetes creators' Mecatl harness splits the coding-agent loop from its tools and session state

Kubernetes creators Craig McLuckie and Joe Beda's startup Stacklok is building Mecatl, a coding-agent harness designed to run in the cloud. It pulls tool calls, session state and memory out of the agent loop so each can be put into systems an operator can manage.

Publishers:latent.space

Reality

Evidence35
Adoption
Not scored
Hype gap+20
Incentives75
Confidence40
buildOne report1 publisher

Pi's v0.99.0 adds MCP by running tool calls in a sandbox outside the model's context

Pi shipped MCP in v0.99.0 using a sandbox that keeps tool schemas out of context, where three servers took 143,000 of Perplexity's 200,000 tokens. Other harnesses can copy the design if they are willing to host an interpreter that runs model-written code.

Publishers:dev.to

Reality

Evidence32
Adoption
Not scored
Hype gap+35
Incentives
Not scored
Confidence35
buildOne report1 publisher

A static denylist stopped one more prompt injection than no protection in a coding-agent study

Bouras, Dai and Mechtaev found a static denylist let 46 of 75 prompt injections execute in a coding agent, against 3 under preflight-scoped capabilities. A same-day Google report of malware stealing OIDC tokens from GitHub Actions runners puts the outer limit on an agent in the CI job's permissions.

Publishers:dev.to

Reality

Evidence45
Adoption
Not scored
Hype gap+10
Incentives
Not scored
Confidence50
buildOne report1 publisher

Pi 1.0 runs MCP tools inside a sandbox so their output stays out of context

Earendil's Pi 1.0 coding agent ships native MCP support after more than a year of criticising the protocol, running tool calls inside a WASM sandbox. The team credits a more stateless July 2026 spec and a sandbox it already needed for non-LLM models, so MCP support was a small addition.

Publishers:dev.to

Reality

Evidence40
Adoption35
Hype gap+20
Incentives40
Confidence40
buildOne report1 publisher

Pi Durable checkpoints every agent step so crashed runs resume from storage

Pi Durable ports the Pi agent harness to TypeScript and checkpoints every step to one of three stores, so crashed agents resume where they stopped. Whether it can replace hand-built resume code depends on how it treats a tool call cut off mid-flight.

Publishers:latent.space

Reality

Evidence28
Adoption
Not scored
Hype gap+30
Incentives
Not scored
Confidence35
buildOne report1 publisher

Pi adds MCP through a JavaScript sandbox that skips tool-schema preloading

Pi made MCP a core feature in version 0.99.0 on September 29, 2026, after its creator Mario Zechner spent over a year arguing against the protocol. In Pi, models call MCP tools from JavaScript in a QuickJS sandbox, so tool schemas are not preloaded into context.

Publishers:dev.to

Reality

Evidence40
Adoption
Not scored
Hype gap+30
Incentives
Not scored
Confidence45
buildConfirmed3 publishers

On-prem agents are now a purchasing decision: Perplexity's local Computer needs $4,800 of Nvidia

Portable Computer's harness is real engineering. The bill of materials starts at a $4,800 DGX Spark or a 24GB RTX card, which turns local agents into a hardware order.

Perspective Coverage

3 publishers
Builder
Builder 33%
Operator
Operator 40%
Investor
Investor 27%

Reality

Evidence55
Adoption
Not scored
Hype gap+20
Incentives70
Confidence60
buildOne report1 publisher

An SSH key upload scoring 0.02 survives this agent gate once the threshold hits 0.99

Jev answers propositions with calibrated probabilities, and the gate around it mirrors every threshold on 0.5, so a two-point tightening of the secret-egress cutoff deleted two thirds of the band that catches exfiltration.

Publishers:dev.to

Reality

Evidence38
Adoption8
Hype gap+10
Incentives40
Confidence45
buildOne report1 publisher

A nullable parentId buys Pi the entire session tree for one field

Codex made the session Item a wire type and Pi gave every entry a nullable parentId, while Claude Code squeezes the transcript in five stages. The three designs diverge on what stays addressable after compaction.

Publishers:dev.to

Reality

Evidence58
Adoption30
Hype gap+12
Incentives30
Confidence60
buildOne report1 publisher

Pi runs the model's shell commands with the permissions of whoever launched it

A seven-part video walkthrough is packaging the Flask creator's Pi workflow for a general audience. The harness itself leaves isolation to the operator, and Earendil's own documentation says so plainly.

Publishers:runtimewire.com

Reality

Evidence66
Adoption18
Hype gap+22
Incentives66
Confidence55
buildOne report1 publisher

Seven harnesses took the transcript. None of them took the approvals.

A frozen Claude Code session moved cleanly into seven rival agent formats. The approval and audit machinery wrapped around it moved into none of them.

Publishers:dev.to

Reality

Evidence58
Adoption20
Hype gap+12
Incentives38
Confidence55
buildOne report1 publisher

Pi ships four tools and no sandbox, so the guardrails are on your build list

A review of v0.84.2 is blunt: no built-in sandbox, no permission layer. Three of the four default tools change the machine, and whatever sits between them and your repo is code you own.

Publishers:dev.to

Reality

Evidence34
Adoption16
Hype gap−8
Incentives44
Confidence38
buildOne report1 publisher

Hash-anchored edits turn a lost race into an error, and that contract outranks the tool count

Oh-My-Pi makes a write conditional on the target lines still hashing the same, so the loser of a race gets an error instead of a wrecked file. The guarantee is paid for in retries.

Publishers:dev.to

Reality

Evidence26
Adoption14
Hype gap+34
Incentives
Not scored
Confidence31
buildOne report1 publisher

Zalando's durable agentic engineering win was a proxy, not a model

A 2.5-year retrospective spanning more than 250 engineering teams credits a LiteLLM-based API proxy, stood up in January 2024, for metering adoption and forcing client upgrades.

Publishers:engineering.zalando.com

Reality

Evidence46
Adoption64
Hype gap−12
Incentives52
Confidence55