This repository is a standalone toolkit for the Pi coding agent, and it stays one.
Do not mix it up with the agent fleet. Every CLONE FRAME / iCLONE agent and every iNFT forged from
inft-i01runs on the Hermes Agent (Nous Research, MIT) — Pi was retired as the substrate on 2026-08-29.
Substrate Where The iNFT / agent fleet Hermes Agent (Nous Research) inft-i01and every repo forged from itThis repo Pi coding agent (Earendil) standalone tool, not part of that stack Nothing here was rewritten into a Hermes tool, because it is not one — it extends Pi, and pretending otherwise would be a lie. If you are looking for the iNFT agents, you are in the wrong repository.
Every Pi extension as an in-session slash command. Agent teams, pipelines, subagents,
peer-to-peer coms and safety rails — activated and deactivated with a single command,
inside the session you are already in. No relaunch. No -e flags. No restart.
Built on the Pi coding agent · engines from pi-vs-claude-code by IndyDevDan (MIT).
Pi extensions are powerful, but they load at startup:
# before — want the agent team? leave your session and start over:
pi -e extensions/agent-team.ts -e extensions/theme-cycler.ts
Change your mind mid-session — need subagents instead, or want the safety rails off — and it's quit, retype flags, relaunch, lose your flow.
agentic-teams-pi turns every extension into a switch:
# after — inside the session you already have:
/pi-team → team dispatcher armed, grid dashboard live
/pi-team → same command again: disarmed, gone
/teams → see every team, chain and agent you have, right now
/pi-off → everything off, plain pi again
Under the hood, each command writes the engine into ~/.pi/agent/settings.json and calls
Pi's extension reload — so the engine's commands, tools and widgets appear (or disappear)
live in the current session, and the choice persists for future sessions until you
toggle it back.
Live boards for everything: expert grids answering in parallel, dispatcher cards animating
on every dispatch_agent, pipeline steps lighting up, subagent cards streaming their work:
| Path | For | You get |
|---|---|---|
| A. Slash commands — default | everyone | Every feature as an in-session switch: /pi-team on, /pi-team off. Installed once, available in every pi session. |
| B. Raw extensions — pro | users who want the classic -e workflow |
The 18 engines, launched and stacked by hand, no slash layer. Full reference: docs/EXTENSIONS.md |
🤖 Installing with an AI agent? Point it at this repo and tell it what you want — it will find AGENTS.md, which tells any LLM exactly how to install and verify either path. Saying nothing special gets you Path A.
Run inside this repo (project settings load the slash layer for sessions started here):
git clone https://github.com/devclone20/agentic-teams-pi.git
cd agentic-teams-pi
piToggles are saved to your global
~/.pi/agent/settings.json— that is what makes them survive into your next session./pi-off(or./uninstall.sh) cleans them out. Two engines (/pi-team,/pi-chain) read YAML and need a one-timebun install(ornpm install) in the clone when used in-repo without the global install.
Or install globally (commands available in every pi session, in any directory):
./install.sh # copies engines to ~/.pi/agent/agentic-teams + wires settings.json
pi # anywhereA fresh install also enables the visual starter set — team grid, expert grid, chains
board, subagent cards and the minimal footer — so your first session already shows the
boards (skip it with n at the prompt, or trim it any time: each feature's own command
turns it off, /pi-off turns off everything). Re-installs never touch your toggles.
Then type:
/teams
and the roster board appears — every team, chain and agent you own, plus how to arm them.
git clone https://github.com/devclone20/agentic-teams-pi.git
cd agentic-teams-pi
pi -e extensions/agent-team.ts # one engine
pi -e extensions/minimal.ts -e extensions/tool-counter-widget.ts # stacked
just ext-agent-team # or the ready-made combos"Off" is a launch without the flag; a persistent set is engine paths added to
~/.pi/agent/settings.json → extensions (exactly what the slash layer automates).
Per-engine flags, config files and env vars: docs/EXTENSIONS.md.
Same command = on and off. ● needs a config file (a starter set ships in .pi/agents/).
The always-current version of this table lives in docs/COMMANDS.md and
inside any session via /commands-pi.
| Command | What it does | What it shows |
|---|---|---|
/teams |
The roster board — every team, chain and agent available right now | Teams with members, chains with step counts, agents as slash commands, engine state |
/pi-list |
Status board of every feature | Grouped board with ●/○ dots: what's on, what's available |
/commands-pi |
Full in-session reference | Every command + what it unlocks + personas + prompt templates |
/pi-themes [name] |
Instant theme switch (autocompletes names) | Theme picker, or direct switch |
/clear |
Wipe every message from the screen — fresh session | A clean terminal with only your agent boards and widgets |
/pi-off |
Panic button: deactivate every engine | Summary of what was turned off |
| Command | What it does | What you see when active |
|---|---|---|
/pi-team ● |
Loads the team dispatcher engine (dormant until you pick a team — then the primary agent delegates via dispatch_agent) |
Live team grid above the editor; /agents-team picks & arms the team, /agents-list, /agents-grid <1-6> |
/pi-chain ● |
Loads the pipeline engine (each step's output feeds the next agent: plan → build → review) |
Pipeline widget with step cards; /chain picks & arms the pipeline, /chain-list shows all |
/pi-sub |
Background subagents: offload tasks while you keep working | One live streaming widget per subagent; /sub <task>, /subcont <id> <prompt>, /subrm <n>, /subclear |
/pi-experts ● |
Meta-agent experts: parallel Pi-framework specialists answer research questions | Expert grid with per-expert progress; /experts, /experts-grid <1-5>, query_experts tool |
| Command | What it does | What you see when active |
|---|---|---|
/pi-tilldone |
Task discipline: tools are blocked until a task list exists and one task is in progress | Persistent task list in the footer with live progress; /tilldone opens the overlay |
/pi-system |
Switch the system prompt to any discovered agent persona | Persona picker; active persona in the status line; /system |
/pi-cross |
Registers commands & skills from .claude/, .gemini/, .codex/ dirs |
Each discovered item becomes /name or /skill:name |
/pi-replay |
Session timeline: scrollable replay of everything that happened | Full-screen overlay: every user/assistant/tool step, timestamped; /replay |
/pi-gate |
Purpose gate: declare the session's single purpose before any work | Blocking prompt at activation, then a persistent purpose banner |
| Command | What it does | What you see when active |
|---|---|---|
/pi-coms |
Peer-to-peer agents, same machine (Unix sockets): two equal Pi agents talking, no hierarchy | Live peer-pool widget; coms_list / coms_send / coms_get / coms_await tools |
/pi-coms-net |
Same idea across machines via an HTTP/SSE hub | Peer pool across the network; coms_net_* tools; hub: bun scripts/coms-net-server.ts |
The two coms engines share CLI flags and cannot run together — turning one on automatically turns the other off.
| Command | What it does | What you see when active |
|---|---|---|
/pi-damage strict ● |
Blocks dangerous bash & file access (from damage-control-rules.yaml) and aborts the turn |
🛡️ shield in the status line; blocked calls reported |
/pi-damage continue ● |
Same rules, but the agent keeps working with actionable feedback | Same shield; the turn adapts instead of dying |
/pi-damage off |
Rails off | — |
| Command | What it does | What you see when active |
|---|---|---|
/pi-focus |
Distraction-free mode — again to restore | Nothing. That's the point |
/pi-minimal |
Compact footer | deepseek-v4 [###-------] 30% |
/pi-counter footer |
Rich two-line stats footer | Model, context, tokens, cost, branch + per-tool tally |
/pi-counter widget |
Tool tally as a colored widget above the editor | Tools (12): bash 5 · read 4 · edit 3 |
/pi-cycler |
Theme hotkeys + /theme picker |
🎨 theme in status line; ctrl+shift+x / ctrl+q to cycle (hotkeys are terminal-dependent on macOS — see docs/RESERVED_KEYS.md) |
slash/agent-slash.ts gives every agent persona its own slash command — the fastest
way to fire a specialist at a task:
| You type | What happens |
|---|---|
/scout map the auth flow |
The scout persona runs as a one-shot pi subprocess; its report streams back into your conversation |
/builder implement the fix |
Same, with the builder persona |
/cc-<name> <task> |
Personas from .claude/agents/ — prefixed cc- so they never collide |
/agents-slash-list |
Everything registered, grouped by source |
The starter roster: scout · planner · builder · reviewer · documenter ·
red-team · plan-reviewer · bowser (browser automation via the bundled bowser
skill + bash) + ten Pi-framework experts.
One switchboard, four engine families — a command flips the wiring and reloads in place:
%%{init: {'theme':'base','themeVariables':{'primaryColor':'#1d1830','primaryTextColor':'#e8e3f5','primaryBorderColor':'#ff7edb','lineColor':'#8a7fa8','fontFamily':'monospace','clusterBkg':'#16121f','clusterBorder':'#2a2139','edgeLabelBackground':'#16121f'}}}%%
flowchart LR
U(["you type /pi-team"]):::you --> M["slash/agentic-teams.ts<br/>the switchboard"]:::master
M -->|"writes engine path"| S[("settings.json")]:::store
M -->|"ctx.reload()"| R{{"live reload<br/>same session · context kept"}}:::reload
S -.-> R
R --> O["🎛 Orchestration<br/>team · chain · sub · experts"]:::orch
R --> C["📡 Communication<br/>coms · coms-net"]:::coms
R --> SA["🛡 Safety<br/>damage strict / continue"]:::safe
R --> I["🎨 Interface & Workflow<br/>footer · themes · replay · gate"]:::iface
classDef you fill:#241a2e,stroke:#fede5d,color:#fede5d
classDef master fill:#2a1630,stroke:#ff7edb,color:#ff7edb
classDef store fill:#1d1830,stroke:#8a7fa8,color:#c8bfe0
classDef reload fill:#152530,stroke:#36f9f6,color:#36f9f6
classDef orch fill:#1d1830,stroke:#6a7bdb,color:#9fb0ff
classDef coms fill:#152530,stroke:#36f9f6,color:#36f9f6
classDef safe fill:#241a22,stroke:#f97e72,color:#f9a292
classDef iface fill:#182b22,stroke:#72f1b8,color:#72f1b8
And the full life of a switch — on, off, on again, always the same command:
%%{init: {'theme':'base','themeVariables':{'actorBkg':'#1d1830','actorTextColor':'#e8e3f5','actorBorder':'#ff7edb','actorLineColor':'#4a3f66','signalColor':'#8a7fa8','signalTextColor':'#c8bfe0','noteBkgColor':'#241a2e','noteTextColor':'#fede5d','noteBorderColor':'#fede5d','fontFamily':'monospace'}}}%%
sequenceDiagram
participant You
participant Session as pi session
participant Master as agentic-teams.ts
participant Settings as ~/.pi/agent/settings.json
You->>Session: /pi-team
Session->>Master: run command handler
Master->>Settings: add extensions/agent-team.ts
Master->>Session: ctx.reload() (same session, in place)
Session-->>You: ● Agent Team ON — /agents-team /agents-list /agents-grid ready
You->>Session: /pi-team (later)
Master->>Settings: remove agent-team.ts
Master->>Session: ctx.reload()
Session-->>You: ○ Agent Team OFF — same command turns it back on
Three properties fall out of this design:
- Symmetric — one command is the whole lifecycle: on, off, on again.
- Persistent — toggles survive into your next session until you change them.
- Collision-proof — engines are matched by filename, so a copy loaded from another directory is recognised and replaced, never doubled.
Activation reloads the extension runtime in place. Your conversation and context are untouched, but live widgets (running subagents, chains in flight) reset — arm your engines before launching long-running work.
Teams and chains are plain YAML in .pi/agents/ (project) or ~/.pi/agent/agents/ (global):
# .pi/agents/teams.yaml — /agents-team picks one of these
frontend: [planner, builder, bowser]
quality: [reviewer, red-team]
docs: [scout, documenter]# .pi/agents/agent-chain.yaml — /chain runs one of these
plan-build-review:
steps:
- agent: planner
prompt: "Plan this work: $INPUT"
- agent: builder
prompt: "Implement exactly this plan: $INPUT"
- agent: reviewer
prompt: "Review the result: $INPUT (original ask: $ORIGINAL)"Each persona is a Markdown file with frontmatter (name, description, tools) and the
system prompt as body. Drop a new .md in the folder — it becomes a team member and
a /name slash command.
# terminal 1 # terminal 2
pi pi
/pi-coms /pi-comsEach side sees a live pool of peers. Ask one agent to coms_send a prompt to the other
and coms_await the reply — two equal agents, prompt → response → prompt, no hierarchy.
For different machines, start the hub (bun scripts/coms-net-server.ts) and use
/pi-coms-net instead.
Also in the tree: scripts/coms-net-server.ts (the HTTP/SSE hub for /pi-coms-net),
scripts/gen-commands.ts (regenerates docs/COMMANDS.md from the REGISTRY) and the
justfile (just pi · just smoke · just rpc-check · just gen-docs · just ext-*).
| Tool | Needed for | Install |
|---|---|---|
| pi ≥ 0.82 | everything | npm i -g @earendil-works/pi-coding-agent |
| node | install.sh settings edit |
ships with pi's install |
| bun | tests, docs generation, coms-net hub (optional otherwise) | bun.sh |
| just | task runner (optional) | brew install just |
API keys: Pi does not auto-load .env — copy .env.sample to .env and source .env
before launching, or configure providers with /login inside pi.
just smoke # static: every module loads, zero collisions, every engine covered
just rpc-check # live: boots pi in RPC mode in a throwaway HOME, flips real toggles- IndyDevDan — the 19 engines began life in pi-vs-claude-code (MIT). Watch Pi Coding Agent: The Only Claude Code Competitor.
- Mario Zechner — the Pi coding agent itself.
- Alex Rider (devclone20) — the in-session slash-command layer this repo exists for.
MIT — see LICENSE.