pi vs Claude Code vs OpenCode: Picking Your Agentic CLI

Last updated: August 23, 2026

This is part three of our pi series. Part one covered pi's minimal toolkit architecture and part two was a hands-on tour of run modes and session trees. Both ended the same way: fine, but would you make it your main agent? That question is only answerable against the incumbents. So this post compares pi against Claude Code, OpenCode, and Codex CLI on the things that decide a daily-driver choice: what each tool refuses to do, how far its model freedom goes, whether your subscription actually carries over, what it phones home, and what it costs when you lean on it hard.

Everything below was checked against official documentation, primary repositories, and the GitHub API on August 23, 2026. Where a capability is genuinely not documented anywhere official, the matrix says so rather than guessing.

Official Sources#

Where Each Harness Stands on August 23, 2026#

pi: the kit that refuses to be a product#

pi is an MIT-licensed TypeScript monorepo from Earendil holding roughly 95,900 stars and 11,900 forks (GitHub API), at version v0.84.2 released August 14, 2026 after roughly one release per day since December 2025. Out of the box the model gets four tools (read, write, edit, bash), and the project's philosophy section explicitly refuses to build MCP support, sub-agents ("spawn pi instances via tmux"), permission popups, plan mode, built-in to-dos, and background bash. Extensions, skills, and packages fill those gaps if you want them filled - our hands-on guide walks the mechanics. Its session format is the differentiator: JSONL files where every entry points at a parent, so /tree, /fork, and /clone navigate real branches of a conversation (session-format docs). The community's own taxonomy from the August 4 Hacker News thread holds up: "Codex, Claude Code are VS Code, JetBrains. Pi is Neovim" (HN).

Claude Code: the full product#

Claude Code is Anthropic's proprietary agent, shipped as a native binary plus VS Code, JetBrains, desktop, web, and mobile surfaces that all share one engine (overview). Its public GitHub repository shows 142,749 stars but carries no license file - the product itself is closed source (anthropics/claude-code). The orchestration stack is the deepest in the industry: subagents that run in the background by default and can spawn their own subagents (sub-agents, what's new week 24 and 27), agent teams with inter-agent messaging (agent teams), scriptable dynamic workflows (workflows), skills, plugins and marketplaces, hooks, MCP, cloud Routines, and an Agent SDK in TypeScript and Python (Agent SDK). Model coupling is total by design: you run Claude models (Fable, Opus 5, Sonnet 5, Haiku), and the enterprise routes through Amazon Bedrock, Google Cloud's Agent Platform, or Microsoft Foundry change where inference runs, not whose models run (feature availability).

OpenCode: the open-source middle path#

OpenCode is MIT-licensed, built by Anomaly (the SST team), and now the largest open-source agent in the category: 200,642 stars and 25,944 forks as of August 23, 2026 (GitHub API), up from the 160,000 we recorded in June in our OpenCode developer guide. Latest release v1.18.21 landed August 21, 2026. It ships a TUI, desktop app, and IDE extension (docs), brings your own keys for 75+ providers, and layers first-party billing on top: Zen, a curated pay-per-token model marketplace with zero markups, and Go, a $10-per-month subscription ($5 the first month) targeting roughly 6x usage value with documented caps of $12 per five hours, $30 weekly, and $60 monthly. Plan mode and build mode are built in via Tab, sessions live locally, and the extension surface covers plugins, custom tools, commands, formatters, themes, LSP servers, and MCP (configure section).

Codex CLI: the fleet tool with fresh models#

Codex CLI is OpenAI's open-source Rust agent: Apache-2.0, 115,076 stars (GitHub API), latest release rust-v0.149.0 on August 20, 2026. It matters right now because of what it runs: the GPT-5.6 family went generally available on July 9, 2026 with three durable tiers - Sol for complex work, Terra as the everyday workhorse, Luna for high-volume tasks (models doc; announcement). Plan availability is tiered: Free and Go ChatGPT users get Terra only, while Plus, Pro, Business, and Enterprise choose among all three, including the max reasoning effort in Codex (announcement). Two August wrinkles worth knowing before you standardize: GPT-5.4 and gpt-5.3-codex are deprecated under ChatGPT sign-in with retirement at the end of August (migration guide), and the documented 1M-token context window is currently capped much lower inside Codex clients - Sol at 272K in the bundled catalog, a gap tracked and closed-without-fix in issue #38917, with follow-up reports putting Terra and Luna at 872K (issue #39144).

One more entrant deserves a sentence: Vercel's fx, a Zig-based "Unix like coding agent" created August 11, 2026, already at 2,225 stars (vercel-labs/fx). The minimalist-harness lane pi opened is now crowded enough that new builds target it explicitly. We have not tested fx and exclude it from the matrix.

Capability Matrix#

Every cell links or cites its basis. "Not documented" means we could not find official documentation as of August 23, 2026.

Five Rows That Decide the Purchase#

Model freedom vs model depth. pi will happily run Fable 5, then GPT-5.6 Terra, then a local llama.cpp model in one session, converting thinking traces across providers as it goes (origin writeup). OpenCode gets you the same breadth through provider keys but without pi's mid-session handoff polish. Claude Code and Codex are walled gardens on purpose - and both walls are currently defensible, because Opus 5 (default since July) and GPT-5.6 Sol sit at the top of independent coding indexes (Artificial Analysis, via OpenAI). You are trading optionality for a tuned loop either way.

The subscription-reuse trap. This is the row most people get wrong. pi accepts your Claude credentials through OAuth, but the providers documentation states plainly that third-party harness usage draws from extra usage billed per token, not against Claude plan limits. Your Max subscription does not follow you into pi. The ChatGPT route is different: OpenAI officially endorses pi's Codex-based login through its Codex for OSS program. Meanwhile Claude Code and Codex treat their subscriptions as the whole point, and OpenCode sells you its own cheaper rails instead.

Orchestration is a spectrum, not a checkbox. Claude Code ships backgrounded nested subagents, teams with messaging, and rerunnable workflows (agents). OpenCode gives you structured primary/subagent definitions. Codex pushes parallelism outward into cloud tasks and headless fleets rather than in-session spawning. pi refuses the entire category and hands you tmux - which, as we found in the Herdr comparison, is more observable but entirely on you.

Session trees are becoming table stakes, three ways. pi's branching JSONL remains the purest expression: alternatives live in one file you can parse in twenty lines (session format). Claude Code responded with branchable, fork-default sessions and checkpoint rewinds (sessions, week 33) - closed format, capable UI. OpenCode and Codex CLI have nothing comparable documented today.

Telemetry posture is now a selection criterion. Anthropic documents an unusually detailed telemetry surface: metrics on by default, error reporting on for subscription sign-ins, a consumer training toggle that flips retention between 30 days and 5 years, and a WebFetch hostname check that always runs (data usage). Everything has named opt-outs, which is better than most, but the defaults are chatty. pi keeps everything in local plain-text JSONL by construction. For proprietary code under compliance pressure, this row alone can make the decision.

Verdict Profiles#

Pick pi if...#

You are a developer who wants to own the harness. Concretely: you want one agent loop across any provider including local models, sessions as inspectable files, scripting surfaces from one-liners to an embedded SDK, and you accept assembling permissions (containers), sub-agents (tmux or packages), and todos yourself. The Neovim tax is real - expect to spend an evening on extensions and to pin versions on a pre-1.0 project shipping near-daily (releases). The payoff is a tool nobody can take away from you and the lowest measured cost per task in its benchmark class.

Pick Claude Code if...#

Your work is long-horizon surgery on one large codebase, and you want the machine to handle orchestration, not configure it. Skills encode your team's repeated workflows, agent teams fan out coordinated work, routines automate the boring loops, and every surface from phone to CI shares one engine (overview). You pay for that with model lock-in, a shared quota pool that chat competes with (pricing FAQ), and the most instrumented telemetry defaults in this comparison (data usage).

Pick OpenCode if...#

You want an open-source, model-agnostic main agent with product polish and predictable costs. It sits deliberately between pi and Claude Code: more batteries than pi (plan mode, undo, LSP integration, MCP), more freedom than Claude Code (75+ providers, MIT), and first-party pricing that undersells both - $10 a month on Go for capped open-model usage, or Zen at token-plus-nothing (Go). Our developer guide covers setup end to end. Choose it over pi when you want features preassembled; choose it over Claude Code when vendor or price lock-in is the deciding factor.

Pick Codex CLI if...#

You already pay for ChatGPT Plus or above and your bottleneck is supervising many concurrent agents rather than depth in one session. Worktree isolation, delegated cloud tasks, and codex exec make it the strongest fan-out harness here (our field notes), and the GPT-5.6 price cuts made its quota stretch further in July and again on August 21 (price post). Verify context-window behavior against your workload before committing - the client caps are not yet what the marketing pages imply (issue #38917).

Who should NOT leave Claude Code#

Be honest about fit before migrating. Stay put if: your team runs on managed settings, analytics, gateway spend limits, or zero-data-retention agreements - none of which the open tools document equivalents for (admin setup); your workflows are encoded in skills and hooks that assume Claude-model behavior; your repos are large enough that long-context coherence is the binding constraint; or you simply cannot spend configuration hours, because pi's flexibility is purchased exactly there. A harness migration is a week of friction for a capability delta you may not need - the correct move for most Claude Code shops is running pi or OpenCode as a second, cheap, scripted lane before considering a divorce.

FAQ#

Is pi ready to be a main harness or still a hobby?#

Ready, with conditions. It drives real production work for its ecosystem - distributions like LazyPi and oh-my-pi exist because people run it daily (HN minimalism thread) - but it is pre-1.0, moves at nearly a release a day, and expects you to assemble safety and orchestration. If you want software, not a kit, it will feel unfinished.

Can I reuse my Claude Max or ChatGPT subscription in pi?#

ChatGPT Plus/Pro: yes, officially endorsed via the Codex OSS program. Claude Pro/Max: technically yes through /login, but the docs state usage draws from extra usage billed per token, not your plan allowance (providers doc). Budget accordingly - this surprises people at invoice time.

Which tool is cheapest for heavy daily use?#

Order of magnitude, cheapest first for comparable volume: pi on discount API providers or local models, OpenCode Go at $10/month with published caps, Codex on post-cut Luna/Terra rates, then Claude Code Max tiers (matrix above). But raw token price is not cost per task - pi's lean context repeatedly measured cheapest per completed task in the benchmark we analyzed (cost post), and a harness that burns half the tokens can beat a bigger subscription.

Does anything else match pi's session trees?#

Not yet in open form. Claude Code now branches and forks sessions with checkpoints, which covers the interactive workflow (sessions), but its transcript format is not a documented, user-parseable tree. OpenCode and Codex CLI document resume but not alternative-branch exploration. If forking experiments is how you think, pi is still the reference implementation.

What about Vercel's fx?#

A new minimalist entry: Zig, Unix-style composition, repository created August 11, 2026, 2,225 stars by August 23 (vercel-labs/fx). It validates the lane pi opened but is too young for this matrix - no track record, no extension ecosystem, no pricing story yet. Watch it; do not standardize on it.

Is Claude Code's telemetry bad enough to switch over?#

That depends on your data. The behaviors are documented and individually opt-out-able: metrics, error reports on subscription sign-ins, feedback retention measured in years, and an always-on hostname preflight for WebFetch (data usage). Commercial terms exclude your code from training, and ZDR exists for qualified enterprises. If even documented defaults are unacceptable for your repos, pi's local-JSONL posture or OpenCode's local-first sessions are the cleaner answers - just lose the features you traded away.

Which should a team standardize on?#

For regulated or larger engineering orgs: Claude Code, because team controls, deployment plumbing, and compliance paperwork are actual products there (admin setup). For startups and open-source teams: OpenCode or Codex on price and flexibility. Standardize on pi only if at least one engineer owns the harness as part of their job - it rewards an owner and punishes a vacuum.

Can extensions really close pi's feature gaps?#

Mostly, and that is the design bet. The README's own list includes community-built sub-agents, plan mode, permission gates, and MCP bridges (customization), and our hands-on guide verified working examples from sandbox policies to Herdr worker fleets. What extensions do not close well today is turnkey sandboxing - the community calls third-party options half-baked (HN) - so containerize properly rather than improvising.

---

That closes the series. Start with pi's architecture to understand why the minimal bet exists, work through run modes and session trees to feel the difference in your own terminal, then use this post to decide whether pi becomes your main agent, your second agent, or a philosophy you borrow ideas from. When you do run multiple agents side by side, Herdr vs pi vs tmux covers the layer above them all, and the broader four-way harness shoot-out remains the map if Cursor enters your picture.