Pi coding agent alternatives: choose by exit reason and migration cost

You do not have to leave the Pi coding agent for lifecycle reasons: on September 21, 2026 the GitHub API reported earendil-works/pi as MIT-licensed and not archived, with release v0.87.0 published the same day. The real reasons people move are design choices. Until a September 22, 2026 docs refresh, Pi's own docs said it "intentionally does not include built-in MCP, sub-agents, permission popups, plan mode, to-dos, or background bash" and told you to "build or install those workflows as extensions or packages, or use external tools such as containers and tmux". The refresh removed that sentence, and v0.99.0 on September 29, 2026 shipped one of the six — MCP is now a built-in extension that loads by default. At v0.99.2 the other five are still not built in. So the useful question is not which agent is better, but which of those five you wanted built in — and what the move costs you out of ~/.pi/agent.

One disambiguation, because the search results mix at least five things. This page is about the terminal coding agent at pi.dev, repo earendil-works/pi, npm @earendil-works/pi-coding-agent. It is not Pi by Inflection AI at pi.ai, a different product entirely; not the Pi Network coin; not Raspberry Pi; and not the constant. The old shittycodingagent.ai domain now redirects to pi.dev, and badlogic/pi-mono resolves to the current repo, so old links land in the right place without telling you they moved.

Match the reason you are leaving to a destination

Each row below is a reason that has appeared in Pi's own documentation as a deliberate omission or boundary, paired with a project that documents the thing instead. Nothing here was runtime tested.

Two boundaries worth naming before you pick. Pi's security page, as rewritten in its September 22, 2026 docs refresh, says plainly that Pi "can read, change, and execute files with the permissions of the account that started it, and it does not ask for approval before every tool call", and it lists the "lack of a built-in sandbox" among behaviours generally outside its security boundary. Its project-trust prompt is not the missing permission system either: the same page says trust "controls whether Pi loads most settings and resources supplied by a working folder" and "does not limit what tool calls can access or affect". If isolation is your exit reason, an approval-mode client or a container is the answer, not a different terminal agent by default.

Check the status before you commit to a name

Alternatives lists for Pi carry projects in three different states, and one field on GitHub actively misleads. Read release and commit dates, not pushed_at.

Repository facts read from the GitHub API and each project's own site on September 21, 2026. Star counts and push timestamps drift daily and are deliberately left out. Continue is the row to read twice: its pushed_at field shows a recent date because of documentation-only commits, which is how it keeps appearing on current recommendation lists. Kunavo's own Continue alternatives and Aider alternatives pages cover those two projects in their own right.

What actually leaves ~/.pi/agent

This is the part the comparison tables skip, and it is where the real switching cost sits. Pi keeps everything in one directory, and the seven things in it do not share a fate.

Read from Pi's usage, security, models, sessions and providers pages, OpenCode's rules page and Oh My Pi's models documentation, all on September 21, 2026; the auth.json row's resolution order was re-read from Pi's refreshed models page on October 1, 2026, and the mcp.json row was read from its MCP page the same day.

Doing the move without burning the way back

# 1. Change nothing in Pi. Everything below leaves ~/.pi/agent intact.

# 2. Rules travel as-is. Pi already reads the repo's AGENTS.md;
#    so do the destinations that document AGENTS.md support.
cp ~/.pi/agent/AGENTS.md ~/.config/opencode/AGENTS.md   # global rules, OpenCode

# 3. Of the destinations checked here, only Oh My Pi documents a converter,
#    and it only fires from its OWN directory — not from ~/.pi/.
mkdir -p ~/.omp/agent
cp ~/.pi/agent/models.json ~/.omp/agent/models.json     # omp writes models.yml

# 4. Do NOT copy auth.json. Mint credentials in the destination instead.

# 5. Roll back at any time: pi still reads ~/.pi/agent, sessions included.
pi

Two details make the difference between a clean switch and an afternoon of 404s. First, keep the destination on a separate worktree or branch until one complete edit, test and review cycle passes there — Pi keeps reading ~/.pi/agent the whole time, sessions included, so rollback is just running pi again. Second, re-enter your cost and limit metadata in the destination rather than assuming it carries over. Pi's source (v0.99.2) defaults a self-added model's cost block to all zeros, along with a 128,000-token contextWindow and 16,384 maxTokens — so the Pi cost readout you are migrating away from may have been showing $0 the whole time. Whether your destination has the same blind spot is worth checking before you treat its number as a budget. Pi pricing carries the full custom-provider block and its gotchas; this page deliberately does not duplicate it.

If you are weighing a specific pairing rather than a category, Oh My Pi vs Pi covers the fork in detail, Pi vs OpenCode and Pi vs Claude Code cover those two head to head, and OpenCode alternatives runs the same exercise from the other side.

Changing client does not change the token bill

None of these clients charges you for the client itself, and each one can run against whichever model route you configure — that is what the api, baseUrl and key fields in the tables above are for. Several vendors also sell a first-party alternative to your own key (Radius, Zen, Hyper, ClinePass, the OpenHands provider), and Claude Code and Codex CLI additionally sign in against their vendor's own subscription plans, so the figures below describe the bring-your-own-route case only. Either way the exit reason has to be a capability rather than a price: switching agents does not by itself change what the model costs.

The figures below are illustrative token arithmetic, not measured task costs and not a ceiling on a bill. Assume one session that sends 200,000 uncached input tokens and receives 15,000 output tokens, with no cache reads, no cache writes, no image input and no tool charges. Rates render from the live Kunavo catalog per million tokens.

Scale by your own sessions per day before treating any of it as a budget, and keep the two claims separate: the cheapest listed rate and the cheapest way to finish the task are different things, because a smaller model that needs three attempts can cost more than a stronger one that lands the change once. Kunavo's catalog amount is a billing floor rather than a cap — when the upstream reports its charge, the bill is the greater of catalog cost and upstream cost times the applicable markup, as billing explains. The minimum top-up is $10 in prepaid credit, which is a funding minimum, not a task fee or a subscription.

One capability gap to plan around if you are comparing agents with built-in codebase indexing: Kunavo does not serve embeddings. It serves no embedding model, and no text-to-speech or speech-to-text model either, so an indexing step that needs an embedding endpoint has to go to a separate provider.

Kunavo publishes a configuration reference for Pi and for several destinations on this page. A published setup guide is not a compatibility test: none of these clients has been runtime-tested against Kunavo's endpoint, and every protocol claim above was read from each project's documentation rather than run. Start at the Pi integration guide or the Oh My Pi one if the fork is your destination, keep a working route available while you test, and create a Kunavo account when you are ready to price one bounded task against the model you shortlisted.

Frequently asked questions

What was checked here, and how. Pi's and Oh My Pi's repository status and latest releases were read from the GitHub API on September 21, 2026, and Pi's usage, security and models documentation, Oh My Pi's models documentation and OpenCode's rules page were read the same day. After Pi's September 22, 2026 docs refresh, its credential order and api list were re-checked on its models page on October 1, 2026, and the custom-model defaults, which the refreshed docs no longer state, in its source at v0.99.2. The same day, after Pi v0.99.0 shipped built-in MCP, the omitted-feature list was re-checked against its usage, MCP, settings and CLI pages and the source at v0.99.2, the sandbox and project-trust quotes against its rewritten security page, and the api values its source dispatches. The remaining rows — OpenCode, Crush, goose, OpenHands, Cline, Codex CLI, Aider and Continue — were read the same day from the GitHub API and from each project's own linked page: OpenHands' pricing table, Cline's ClinePass page, OpenCode's Zen page, Crush's README and LICENSE.md, goose's provider docs, Codex CLI's config reference and continue.dev. Nothing here was reconciled against a live account on any of those services. Kunavo rates render from the live catalog and every dollar figure here is illustrative token arithmetic. No client on this page has been runtime-tested against Kunavo.