A curated collection of 13 extensions, 1 skill, and 8 themes for
Pi Coding Agent.
Most install with a single pi install npm:<name>.
Browse the catalog · Install everything
Find what you need, copy the install command, reload Pi (/reload).
| Extension | What you get | Install |
|---|---|---|
| claude-code-pi | Claude models (sonnet, opus) via local claude -p |
pi install npm:claude-code-pi |
| grok-pi | Grok CLI models: Grok 4.6/4.5 | pi install npm:grok-pi |
| opencode-pi | Free OpenCode CLI models, no login required | pi install npm:opencode-pi |
| agy-pi | agy CLI models (Gemini, Sonnet, GPT OSS), auto-discovered | pi install npm:agy-pi |
| cursor-pi | Cursor CLI models (auto, Composer, Codex…) via local cursor-agent -p, with install/auth verification |
pi install npm:cursor-pi |
| 9router-pi | 9router gateway models via /v1/models discovery |
pi install npm:9router-pi |
| Extension | What you get | Install |
|---|---|---|
| statusline-pi | Footer: git, PR, context window, tok/s, cost, CPU/MEM | pi install npm:statusline-pi |
| timestamp-pi | Message timestamps + prompt-cache TTL countdown | pi install npm:timestamp-pi |
| subagents-pi | Fleet panel for managed subagents (context, TPS, model) | pi install npm:subagents-pi |
| Extension | What you get | Install |
|---|---|---|
| advisor-pi | advisor tool: strategic guidance from a stronger model |
pi install npm:advisor-pi |
| pi-fusion | delegate tool: a persistent cheap sidekick agent, plus compaction-boundary model routing |
pi -e ./extensions/pi-fusion (not on npm) |
| cache-warm | Opt-in keep-alive pings that avoid prompt-cache misses | pi install npm:cache-warm |
| model-debugger | Log all model requests/responses for provider debugging | pi install npm:model-debugger |
| What | Details | Install |
|---|---|---|
| pi-delegator skill | Delegate approved tasks to a monitored Pi subprocess | repo installer or npm run install-skills from a clone |
neon-green / neon-green-light themes |
Futuristic dark/light pair | repo installer or manual copy (Themes) |
opencode theme |
OpenCode-branded theme | repo installer or manual copy (Themes) |
omarchy theme |
Minimal dark palette inspired by Arch Linux | repo installer or manual copy (Themes) |
Published packages are also listed on pi.dev/packages.
graph LR
A["pi install npm:name"] --> B["~/.pi/agent/npm/"]
C["install.sh --auto"] --> D["~/.pi/agent/<br/>extensions · themes · skills"]
B --> E["/reload"]
D --> E
E --> F["/model — pick bridged providers"]
E --> G["Footer, tools, timestamps active"]
Provider extensions register a Pi provider backed by a local CLI; UI extensions hook the TUI footer and message rendering. Bare extension commands open settings panels in the terminal; provider panels show configuration without editing external files or credentials. Each extension's README documents its controls.
Pick one path:
Single extensions from npm (recommended):
pi install npm:statusline-pipi install npm:advisor-pi # then /reload in PiPin a version, install project-local, or try without installing:
pi install npm:opencode-pi@1.4.0pi install -l npm:statusline-pi # writes .pi/settings.json in this project onlypi -e npm:advisor-pi # current session onlyEverything at once (full collection: extensions + themes + skills):
curl -fsSL https://raw.githubusercontent.com/luongnv89/pi-extensions/main/install.sh | bash -s -- --autoFrom a cloned repo:
git clone https://github.com/luongnv89/pi-extensions ~/.pi/pi-extensions~/.pi/pi-extensions/install.sh --autoManage packages:
pi listpi update --extensionspi remove npm:advisor-piInstaller flags:
| Flag | Effect |
|---|---|
--auto |
Skip prompts, install everything automatically |
--keep |
Keep the cloned repo after installation |
--dry-run |
Show what would be installed without copying |
--repo-url URL |
Use a custom repo URL (default: GitHub) |
--branch BRANCH |
Use a custom branch (default: main) |
Reload Pi after any change: type /reload (or restart).
| statusline-pi — git, cost, CPU/MEM, context zone, tok/s | statusline-pi — wrapped footer on a narrow terminal |
 |
 |
| grok-pi — Grok CLI provider (historical Composer capture) | opencode-pi — free OpenCode models in /model |
 |
 |
| opencode-pi — DeepSeek flash in session | claude-code-pi — claude-code-cli provider |
 |
 |
advisor-pi — strategic advisor tool |
Codex — example Pi session (built-in provider) |
 |
 |
More assets live in assets/ (e.g. statusline-pi-gpt-5-mini-195toks.png, pi-nvidia-kimi-2.6.png).
Quick reference per extension. Full setup guides live in each extension's README.
statusline-pi — custom footer
Replaces Pi's default footer with a compact project statusline.
current-dir │ branch [changed files] PR #x │ remaining context tokens (percentage) context zone │ average response speed │ provider/model
Example:
pi-extensions │ main [2] PR #12 │ 840,037 (84.0%) Plan │ 42.5 tok/s │ openai-codex/gpt-5.5
Git section — groups all git-related status:
- Current branch name
- Number of changed files from
git status --porcelain - Related GitHub PR number (when
gh pr viewresolves one)
Context section — remaining context window as exact tokens plus percentage, followed by the active zone:
840,037 (84.0%) Plan
Zone coloring:
- Plan / Code — success color
- Dump — warning color
- ExDump / Dead — error color
Average response speed — approximate model output speed for the current model/thinking context:
42.5 tok/s
The value averages completed assistant responses, includes the active response while streaming, and remains visible while idle.
Commands:
/statusline-pi # Toggle the custom footer on/off
/statusline-refresh # Force refresh git and PR data
advisor-pi — advisor tool
Registers an advisor tool for strategic planning and course correction.
The executor model can ask a configured advisor model for guidance while keeping
file changes under the executor's control.
Commands:
/advisor-pi status
/advisor-pi enable
/advisor-pi disable
/advisor-pi model <provider>/<model>
/advisor-pi max-uses <number>
/advisor-pi cache <none|short|long>
/advisor-pi reset
Operational notes:
- Each advisor consultation is a separate model call and may add cost.
- Executor streaming pauses while the advisor model responds.
- Cache preferences are passed through where providers support them.
- The advisor has no tools; it only returns strategic guidance.
Full setup: extensions/advisor-pi/README.md
claude-code-pi — claude-code-cli provider
Registers the claude-code-cli provider so Pi can use Claude Code CLI model aliases such as sonnet, opus, and fable. Every model turn spawns the local claude -p command with the selected model; there is no Anthropic SDK, HTTP API, or built-in provider fallback.
pi --provider claude-code-cli --model sonnetCommands: /claude-code-pi status, /claude-code-pi models, /claude-code-pi test, /claude-code-pi help
Full setup: extensions/claude-code-pi/README.md
grok-pi — grok-cli provider
Registers the grok-cli provider so Pi can use the same models as the Grok CLI (currently Grok 4.6 / 4.5). Authenticate with grok login, then pick grok-cli in /model. Reasoning models honor Pi thinking levels (Shift+Tab, /settings, --thinking). Cached metadata supplies context, reasoning, and model pricing when available, including for GROK_PI_MODELS selections.
pi --provider grok-cli --model grok-4.6 --thinking highCommands: /grok-pi status, /grok-pi help
Full setup: extensions/grok-pi/README.md
opencode-pi — opencode-cli provider
Registers the opencode-cli provider so Pi can use free models exposed by the local OpenCode CLI, without opencode auth login.
pi --provider opencode-cli --model opencode/deepseek-v4-flash-freeCommands: /opencode-pi status, /opencode-pi models, /opencode-pi test, /opencode-pi help
Full setup: extensions/opencode-pi/README.md
agy-pi — agy provider
Registers an agy provider so Pi can use models exposed by the
agy CLI — Gemini flash/pro variants,
Claude Sonnet, GPT OSS, and more (extensions/agy-pi/src/index.ts). Models are
auto-discovered from agy models output; effort variants (-high/-medium/-low)
collapse into a single base model, with effort chosen via Pi's thinking-level selector.
Requires the agy CLI on PATH (pip install agy-cli) configured with your API keys.
Environment overrides: AGY_PI_BIN (binary path) and AGY_PI_MODELS
(comma-separated model IDs) (extensions/agy-pi/src/index.ts:508).
pi install npm:agy-pi # or: pi -e ./extensions/agy-piCommands: /agy-pi — show provider status and registered models
cursor-pi — cursor-cli provider
Registers the cursor-cli provider so Pi can use Cursor CLI model ids such as auto, composer-2.5, and gpt-5.3-codex-high. Every model turn spawns the local cursor-agent -p --mode ask --trust command (read-only ask mode); there is no HTTP API or built-in provider fallback. Verifies the Cursor CLI installation and login at session start with fix guidance.
pi --provider cursor-cli --model autoCommands: /cursor-pi status, /cursor-pi verify, /cursor-pi models, /cursor-pi test, /cursor-pi help
Full setup: extensions/cursor-pi/README.md
9router-pi — 9router gateway provider
Registers the 9router provider and discovers the current model catalog from the local gateway. Use /9router-pi refresh after changing enabled models; credentials stay in ~/.pi/agent/models.json.
pi -e ./extensions/9router-piCommands: /9router-pi status, /9router-pi refresh, /9router-pi help
Full setup: extensions/9router-pi/README.md
timestamp-pi — timestamps + cache countdown
Adds a dim timestamp line (with age and message type: user message, ai response, tool call) under every chat message, plus a footer
countdown of the 5-minute prompt-cache TTL (⏳ cache 4:32) that turns green
while warm, yellow under 1 minute, and red once expired
(extensions/timestamp-pi/src/index.ts:177).
Timestamps are stored as custom session entries, so they survive reloads and are rendered TUI-only — never sent to the LLM. Requires Pi >= 0.84.
pi install npm:timestamp-pi # or: pi -e ./extensions/timestamp-piCommands: /timestamp-pi — toggle timestamps and cache countdown
pi-fusion — two-agent sidekick harness
A Pi take on Devin Fusion. Adds a
delegate tool backed by a persistent nested Pi agent session on a cheaper
model: real tools, its own context, reused across delegations, so a follow-up
answers from what the sidekick already learned instead of re-reading the repo.
That persistence is the difference from an advisor-style tool, which re-ships
context uncached on every call.
The main agent is told to delegate and monitor while keeping the plan, the
ambiguous calls, and the final review — and the sidekick is told to flag any
ambiguity it had to resolve rather than guess. The sidekick runs with full coding
tools by default and its nested session has no approval prompts, so it edits
and runs commands unattended; /fusion tools readonly restricts it to
read, grep, find, ls when you would rather apply every change yourself.
Optional /fusion routing on re-evaluates the model assignment at each
compaction — where the cache miss is already paid — stepping the sidekick up
after repeated failures, then the main agent, then back down after a clean
streak. Each delegation reports tokens, real cost, and an estimate of what those
tokens would have cost on the main model.
See extensions/pi-fusion/README.md.
Commands: /fusion status | sidekick | upgrade | frontier | tools | routing | restart | reset
pi -e ./extensions/pi-fusion # from a clone; npm "pi-fusion" is an unrelated packagecache-warm — prompt-cache keep-alive
Keeps the prompt cache alive with hidden keep-alive turns.
Default is off: run /cache-warm on to opt in for the current session.
Keep-alive auto-stops after 30 minutes idle (configurable: /cache-warm duration 1h
or CACHE_WARM_DURATION). Pings use pi.sendMessage() (not sendUserMessage /
completeSimple) and only fire when the session is idle, nothing is queued,
cache activity has already been seen, and remaining TTL is under 60 seconds
(extensions/cache-warm/src/index.ts). The default 15/hour cap supports this
four-minute cadence without forcing an expiry. The 5-minute TTL is an Anthropic
heuristic. Metrics report warm attempts, refreshes, likely avoided misses, and
estimated net USD saved (gross cache discount minus warm-turn spend; N/A when
rates are missing). See extensions/cache-warm/README.md.
pi install npm:cache-warm # or: pi -e ./extensions/cache-warmCommands: /cache-warm on, /cache-warm off, /cache-warm (toggle),
/cache-warm duration [30m|1h|forever], /cache-warm status, /cache-warm metrics
model-debugger — request logging
Logs model requests and responses to ~/.pi/agent/logs/ for debugging
provider interactions.
pi install npm:model-debuggersubagents-pi — subagent fleet panel
Fleet metrics panel for managed subagents (context, TPS, model, thinking);
works with @tintinweb/pi-subagents.
pi -e ./extensions/subagents-pipi-delegator — delegation skill
An agent skill that lets a main AI agent delegate a clear,
approved task to a separate Pi process. It starts by checking available Pi models,
prefers free opencode-cli models by default, saves a reusable default model,
streams progress, and reports duration/token/cost metrics when Pi exposes them.
python3 ~/.pi/agent/skills/pi-delegator/scripts/pi_delegate.py models --prefer-freeUse the skill from Pi as /skill:pi-delegator, or let Pi auto-load it when you
ask to delegate work to a separate Pi instance.
Themes are automatically discovered from ~/.pi/agent/themes/.
| Theme | Style |
|---|---|
neon-green |
Futuristic dark |
neon-green-light |
Softer light variant |
opencode |
OpenCode-branded |
omarchy |
Minimal dark palette inspired by Arch Linux |
aura |
Aura theme's purple/mint palette |
synthwave-84 |
Synthwave '84 neon |
zed-dark / zed-light |
Port of Zed editor's default One Dark/One Light |
Manual install from a clone:
cp ~/.pi/pi-extensions/themes/*.json ~/.pi/agent/themes/Select a theme from Pi's /settings, then reload if needed.
cd ~/.pi/pi-extensions && git pull origin main && ~/.pi/pi-extensions/install.sh --autoThen run /reload in Pi.
pi-extensions/
├── extensions/ # one package per extension (package.json + src/index.ts + README.md)
│ ├── 9router-pi/
│ ├── advisor-pi/
│ ├── agy-pi/
│ ├── cache-warm/
│ ├── claude-code-pi/
│ ├── cursor-pi/
│ ├── grok-pi/
│ ├── model-debugger/
│ ├── opencode-pi/
│ ├── statusline-pi/
│ ├── subagents-pi/
│ └── timestamp-pi/
├── skills/pi-delegator/ # SKILL.md + scripts + references
├── themes/ # aura.json, neon-green.json, neon-green-light.json, omarchy.json, opencode.json, synthwave-84.json, zed-dark.json, zed-light.json
├── scripts/ # packaging guard + npm publish script
├── docs/ # DEVELOPMENT.md, DECISIONS.md
└── install.sh # one-command installer (--auto, --dry-run, ...)
- Contributing Guide — how to add extensions, themes, and submit changes
- Developer Guide — architecture, extension API, theme schema, npm scripts, listing on pi.dev/packages
- Changelog — release history and planned features
- Decisions Log — resolved documentation ambiguities
- Security Policy — how to report vulnerabilities
MIT — see LICENSE for details.