Compare·Minimal coding agent harness
Continuum
Multi-agent workbench
VS
Pi Coding Agent
Minimal coding agent harness

Continuum vs Pi Coding Agent

A harness you can read in an afternoon versus a workbench with a control plane. Weight is the feature; pick deliberately.

Pi is Mario Zechner’s minimal agent harness: four built-in tools, everything else added through a TypeScript extension system, plus skills, prompt templates, themes, and tree-structured session history. It is MIT licensed and near 85,000 GitHub stars as of August 2026, and the monorepo also publishes reusable parts, notably pi-ai (a unified API across 15+ providers and hundreds of models), pi-agent-core, and pi-tui. It moved from badlogic/pi-mono to earendil-works/pi in April 2026 when Zechner joined Earendil, and it stayed minimal through the move. Continuum is the opposite weight class on purpose: a product workbench hosting the official Claude Code and Codex CLIs under your own subscriptions, with git worktrees, a diff and PR review path, native Mac, web, iPhone, and Watch clients, multi-account isolation, live quota gauges, and a local dollar ledger. The useful question is not which is smarter. It is which weight matches the job. See choosing a coding agent.

Download for Mac All comparisons
Workflow Monday path Matrix Decide Money FAQ

Updated 2026-08-03 · Mac stable · Win/Linux desktop beta

Choose Continuum when

You want native clients on four platforms, official Claude and Codex under plans you already pay for, worktree and PR panes, multi-account isolation, and spend by repo, not the smallest possible harness.

Choose Pi Coding Agent when

You want a small, hackable, MIT-licensed harness you fully understand, with a TypeScript extension system, reusable packages, and no product chrome between you and the loop.

Snapshot adjacent job
Dimension Continuum Pi Coding Agent
Job-to-be-done Full multi-provider coding workbench Minimal, hackable coding harness
Surface Mac · web · iPhone · Watch · Win/Linux beta Terminal CLI you extend in TypeScript
Tooling Chat · plan · diff · PR · terminal · artifacts Four built-in tools, rest via extensions
Provider path Official CLIs under your subscriptions 15+ providers via pi-ai, keys or OAuth
Sandboxing Plan gate, git checkpoint, diff review No built-in permission system; use a container
Ops Multi-account · gauges · spawn grid · pairing You compose the loop
Cost view Local JSONL → $ by repo/provider/day Your own metering
Mac stable · web · iPhone · Watch · Win/Linux desktop beta · free app
01

Minimal harness vs full workbench

In Continuum

Product workbench day

09:30

Claude and Codex worktrees open side by side, gauges on the menu bar, iPhone paired before the first standup.

13:00

Plan approve from the phone. The PR pane picks up the customer branch when checks land.

15:30

Second Claude account pinned for the night batch because the primary weekly rail is already past 80%.

17:00

Ledger by repo closes the week. Pi is still one command away for tonight’s local-model experiment.

In Pi Coding Agent

Minimal harness day

09:30

npx @earendil-works/pi-coding-agent against the model backend you configured. Tiny surface, total control.

11:00

Add a twenty-line TypeScript extension for the one tool this task needs, then delete it tomorrow. That is the whole appeal.

13:00

Stay in the loop. No pairing, no daemon, no phone. Run it in a container if it is touching anything you care about.

17:00

Ship or script it. Product chrome never entered the day, by choice, and Pi is the right hammer for a fifteen-line harness experiment.

02

Monday path

Monday if you like small tools but ship multi-provider

Install, connect, first session - steps you can run the same day.

01

Keep Pi installed for local-model spikes and throwaway harness experiments. It starts faster than any workbench will.

02

Install Continuum and attach the Claude and Codex accounts you already pay for, so production tickets ride subscription rails.

03

Spawn Continuum worktree sessions when a task needs isolation, a staged diff, and a PR at the end. See git worktrees.

04

Pair the iPhone when you expect a plan-ready to land while you are away from the desk.

05

Watch the 5h and weekly rails so a Max window never surprises you mid-sprint.

06

If a Pi run needs real filesystem or network reach, put it in a container. Pi ships no permission system and says so plainly.

07

Timebox the choice: throwaway local-model test under five minutes of setup goes to Pi; anything needing phone approve by noon goes to Continuum.

08

Do not try to make Continuum feel like Pi. Different products, deliberately.

How Continuum accepts being heavier

How Continuum accepts being heavier

  • A local daemon, a projection layer, and pairing are real product surface, not a few hundred lines of loop.
  • Official CLIs get Continuum-owned session state: worktree, chat, plan, diff, PR, terminal, artifacts.
  • A git checkpoint is written before every prompt, and plan mode never bypasses approval. Code mode deliberately skips per-tool prompts because the diff pane is the gate.
  • Multi-account isolation uses per-instance config dirs so a second subscription cannot bill the first.
  • The spawn grid and enrolled hosts push work past a single local process onto machines you enrolled.
  • Analytics parse session history locally against a pricing snapshot: dollars by repo, provider, and day, with no telemetry backend.
  • Native apps for Mac, iOS, watchOS, and the web, plus signed distribution, are the cost of multi-device. That is weight with a purpose, not accidental bloat.
03

Adjacent harness matrix

Read this as a weight class table, not a scoreboard. Pi wins every row about thinness and control; Continuum wins every row about devices, accounts, quota, and money. A row marked no is usually a scope decision on one side or the other.

Capability Continuum Pi Coding Agent
Minimal install footprint Full product app + daemon One npm package or curl script
Readable, hackable corefour built-in tools Productized workbench Pi’s design center
TypeScript extension system Skills and slash commands via the hosted CLIs Extensions, skills, templates, themes
Reusable packagespi-ai · pi-agent-core · pi-tui Daemon HTTP/WS API, not an npm library Published and used standalone
Multi-device native apps Mac · iPhone · Watch · web · Win/Linux beta Terminal only
Official Claude / Codex hosts PTY and ACP under your subscription Talks to providers directly, keys or OAuth
Built-in permission system Plan gate + git checkpoint + diff review None by design; containerize instead
Tree-structured session history Sub-chats branch off a parent session First-class in the harness
Worktree + PR panes Chat · plan · diff · PR · terminal · artifacts Your own git workflow
Multi-account Max isolation CLAUDE_CONFIG_DIR / CODEX_HOME isolation One key set per run
Live 5h / weekly gauges Menu-bar rails per provider No plan-window instrument
Local repo $ ledger Local JSONL → $ by repo/provider/day Vendor-neutral telemetry contracts to build on
Local models Through providers and connectors Common Pi use case
Free to start Free app · optional hosted tiers MIT · no hosted service at all
04

Weight as a product choice

01 · Weight

Minimal versus productized.

Pi wins thinness outright: four built-in tools, a core you can read in an afternoon, and packages you can lift into your own agent. Continuum wins multi-device ops, quota rails, multi-account isolation, and cost truth. Neither should apologize for its weight class.

02 · Engines

Official CLIs, productized.

Continuum’s bet is hosting Claude Code and Codex properly, under the subscriptions you already pay for, with version and capability checks so an older host fails honestly instead of silently. Pi talks to providers directly through pi-ai, which is a cleaner contract when you want no CLI in the path at all.

03 · Sandboxing

Say the quiet part on both sides.

Pi ships no built-in permission system for filesystem, process, network, or credential access and points you at containers. Continuum gates with plan mode, a pre-prompt git checkpoint, and per-hunk diff review, and Code mode deliberately skips per-tool prompts. Read both honestly: agent security.

04 · Local models

Admit Pi’s lane.

If the job is a local model and a minimal harness, Pi is usually cleaner. Continuum can reach many providers through connectors, but it is not competing to be the smallest local harness and never will be.

05 · When Continuum wins

The day an ops surface is required.

Phone approve, multi-account Max stacking, worktree isolation across parallel agents, quota rails, a fleet sidebar, and a ledger by repo. See parallel Claude Code sessions for the shape of that day.

05

Coexistence

Stack recipe

How people run both.

Stack recipe: Pi for local-model spikes, scripted one-offs, and anything where you want to own the loop line by line; Continuum for the multi-provider shipping days that need phone approve, worktree isolation, quota rails, and a cost answer. They are not competing for the same slot, and Continuum is not trying to become the smallest harness.

Download Continuum Providers
06

Tiny harness or full workbench

Continuum

Choose Continuum for multi-device shipping ops

  • You want native clients on Mac, web, iPhone, and Watch, not a terminal only.
  • You host official Claude and Codex under subscriptions you already pay for.
  • You need worktrees, a staged diff, and a PR pane for parallel agents.
  • You run more than one account and need config-dir isolation.
  • You want quota rails and per-repo dollars without building analytics.
  • You would rather install a product than assemble and maintain a harness.
Pi Coding Agent

Choose Pi for a harness you own outright

  • You want the smallest coding agent you can fully read.
  • Local models and maximal extensibility are the priority.
  • You want to import pi-ai or pi-tui into something of your own.
  • You are comfortable running agents inside your own container policy.
  • You prefer MIT harness culture to product apps.
  • You are happy scripting your own ops around the loop.
07

What you pay

Continuum

App + your labs

$0 for the app on Mac, web, iPhone, and Watch. Bring your own keys or the Claude and ChatGPT plans you already pay for, and Continuum takes no cut. Optional hosted inference: Plus $25/mo, Max 100 $100/mo, Max 200 $200/mo, Ultra $500/mo.

Pi Coding Agent

Their bill

MIT licensed and free to run, with no hosted service and no paid tier as of August 2026. Your model keys and your own machine are the entire bill, and there is nothing to cancel.

How to compare

Total cost of work

Both are free to install, so compare the ongoing cost of work instead: what your model calls actually cost, and how much of your week goes into maintaining the ops you would otherwise script yourself. Pricing compared.

Pi Coding Agent plans checked 3 August 2026 against their pricing page. Vendors reprice often - if you spot a stale figure, tell us and we will fix it. Continuum never marks up a session it runs under your own login.

Continuum ladder: Free app · Plus $25/mo · Max 100 · Max 200 · Ultra - full pricing. Also analytics, multi-account, devices.

08

Questions

Continuum vs Pi Coding Agent.

Deep dives: docs, providers, sessions.

A minimal MIT-licensed coding agent and agent toolkit from Mario Zechner. It ships four built-in tools and pushes everything else into a TypeScript extension system, with skills, prompt templates, themes, and tree-structured session history. The monorepo also publishes pi-ai, pi-agent-core, and pi-tui as reusable packages.

Yes. Pi is MIT licensed, free to run, and has no hosted service or paid tier as of August 2026. You pay only for the model calls you make.

Either npm install -g --ignore-scripts @earendil-works/pi-coding-agent or the install script at pi.dev. Note the package scope: the project moved from badlogic/pi-mono to earendil-works/pi in April 2026, so older instructions point at the wrong name.

For multi-device, multi-provider coding ops, it is worth considering. For minimal harness work, Pi remains the sharper fit and Continuum is not trying to win that comparison.

Continuum is a commercial product with a free app tier across Mac, web, iPhone, and Watch. Pi is MIT. Different models, and if forkability is a hard requirement, Pi is the answer.

No. Pi documents that it ships no built-in permission system for filesystem, process, network, or credential access, and points you at containerization instead. Continuum gates with plan mode, a git checkpoint before each prompt, and per-hunk diff review, while Code mode deliberately skips per-tool prompts.

Through the providers and connectors Continuum supports, including OpenCode-class paths. For a pure local harness experiment, Pi is usually the simpler route.

Continuum. Per-instance CLAUDE_CONFIG_DIR isolation, per-account gauges, and a fail-closed spawn resolver so a session never silently bills the wrong subscription. Pi does not try to be that product.

Continuum exposes a daemon HTTP and WebSocket API for sessions plus a CLI path, so automation is possible. It is still a product workbench, not a library you import, and Pi stays simpler for pure harness scripting.

Yes, relative to Pi, and that is the honest answer. The weight is the multi-client workbench, the pairing layer, the gauges, and the fleet ops. Choose knowingly rather than being disappointed later.

Further reading

Guide

Git worktrees: the complete guide

One repository, many working directories. The mechanics, the full command surface, and the six things that go wrong.

Read guide → Guide

OpenAI Codex CLI: the complete guide

Install, auth, sandbox modes, config.toml, and the flags that actually change behaviour.

Read guide → Guide

AI coding agents in 2026: the landscape, priced

Four categories, a dozen products, one price table. Pick the category first; the product matters less than people think.

Read guide →
Begin

Run your agents
in Continuum.

Free app. Your subscriptions. Optional hosted inference. Mac stable - Windows and Linux desktop are beta.

Download for Mac See pricing

vendor-neutral · local-first · multi-device