38 KiB
38 KiB
Claude Code vs Pi Agent — Feature Comparison
Pi v0.52.10 vs Claude Code (Feb 2026)
Design Philosophy
| Dimension | Claude Code | Pi Agent | Winner |
|---|---|---|---|
| Core Mantra | "Tool for every engineer" — batteries-included, accessible to all skill levels | "If I don't need it, it won't be built" — minimal, opinionated, built for one engineer's workflow | Both |
| Approach to Features | Ship everything built-in (sub-agents, teams, MCP, plan mode, todos, web search, notebooks, 10+ tools) | Ship the minimum (4 tools, ~200-token prompt). Everything else is opt-in via extensions or bash | Both |
| Safety Philosophy | Safe by default — deny-first permissions, 5 modes, filesystem sandbox, Haiku pre-screening of commands | YOLO by default — no permissions, no sandbox. "Security in coding agents is mostly theater; if it can write and run code, it's game over" | Both |
| System Prompt Trust Model | Extensive guardrails (~10K tokens) — behavioral rules, formatting instructions, safety constraints, tool usage examples | Trust the model (~200 tokens) — "frontier models have been RL-trained up the wazoo, they inherently understand what a coding agent is" | Both |
| Observability | Abstracted — sub-agents are black boxes, compaction happens silently, system prompt not user-visible by default | Full transparency — every token visible, every tool call inspectable, no hidden orchestration, session HTML export | Pi |
| Context Engineering | Managed for you — auto-compaction, sub-agents handle overflow, system decides what enters context | User-controlled — minimal prompt overhead, no hidden injections, "exactly controlling what goes into context yields better outputs" | Pi |
| Extensibility Model | Shell hooks (external processes) + MCP protocol + Skills (markdown prompts) — loosely coupled, config-driven | TypeScript in-process extensions — same runtime, access full session state, block/modify/transform any event | Both |
| Target Audience | Every engineer — beginner-friendly, enterprise-ready, guided workflows, progressive disclosure | Power users — engineers who want control, understand tradeoffs, willing to build their own workflows | Both |
| Multi-Model Stance | Claude-first — optimized for Claude family, gateway workaround for others | Model-agnostic from day one — 324 models, 20+ providers, cross-provider context handoff, "we live in a multi-model world" | Pi |
| Planning Approach | Built-in plan mode — structured explore → plan → code phases, read-only mode, dedicated sub-agents | No plan mode — "just tell the agent to think with you." Write plans to files for persistence, versioning, and cross-session reuse | Both |
| Complexity Budget | Complexity lives in the harness so you don't have to think about it — more magic, less wiring | Complexity lives in your hands — minimal harness, you decide what to add and when. "Three similar lines of code is better than a premature abstraction" | Both |
Cost & Licensing
| Feature | Claude Code | Pi Agent | Winner |
|---|---|---|---|
| Tool License | Proprietary | MIT (open source, fork/embed/self-host) | Pi |
| Subscription Cost | $20-200/mo required or Dedicated Anthropic API Keys | $0 (MIT, BYO API keys) | Pi |
| Cost Tracking | Available via /cost command, customizable via statusline configuration | Real-time $/token/cache display in footer per session and customizable via extensions | Both |
| Cost Optimization | 3 models at 3 price tiers (Opus > Sonnet > Haiku) — single provider | Mix cheap/expensive models per task across any provider, free tiers available | Pi |
| System Prompt Overhead | ~10,000+ tokens | ~200 tokens (more context for actual work) | Pi |
Model & Provider Support
| Feature | Claude Code | Pi Agent | Winner |
|---|---|---|---|
| Official Providers | 4 platforms (Anthropic API, AWS Bedrock, Google Vertex, Foundry) — all serving Claude models | 20+ native (Anthropic, OpenAI, Google, Groq, xAI, OpenRouter, Azure, Bedrock, Vertex, Mistral, MiniMax, Kimi, Cerebras, ZAI, HuggingFace, custom) | Pi |
| Non-Anthropic / Self-Hosted Models | Via ANTHROPIC_BASE_URL gateway — routes to any OpenAI-compatible backend (OpenRouter, LiteLLM, local TGI, vLLM). Functional but unofficial workaround | Native first-class support for all providers + local (Ollama, vLLM, LM Studio via models.json). No proxy needed | Pi |
| Built-in Models | ~6 aliases (opus, sonnet, haiku, opusplan, sonnet[1m], default) mapping to Claude family | 324 (confirmed via ModelRegistry) across all providers | Pi |
| Model Switching Mid-Session | Yes — /model <alias|name> command, --model flag at startup, ANTHROPIC_MODEL env var |
Yes — Ctrl+P cycle, Ctrl+L fuzzy selector, session.setModel() in SDK | Tie |
| OAuth/Subscription Login | Anthropic subscriptions (Pro, Max, Teams, Enterprise) | Claude Pro, ChatGPT Plus, GitHub Copilot, Gemini CLI, Antigravity — all via /login and API keys | Pi |
| Thinking/Effort Levels | 3 effort levels (low/medium/high) on Opus 4.6 via /model slider, env var, or settings |
5 unified levels (off/minimal/low/medium/high) across ALL thinking capable models, Shift+Tab to cycle | Pi |
Agent Harness
| Feature | Claude Code | Pi Agent | Winner |
|---|---|---|---|
| Source Code | Closed source (proprietary) | Open source (MIT license) | Pi |
| System Prompt Size | ~10,000+ tokens (extensive tool descriptions, behavioral rules, safety guardrails) | ~200 tokens (minimal — trusts frontier models to code without hand-holding) | Pi |
| Default Tools | 10+ (Read, Write, Edit, Bash, Glob, Grep, WebSearch, WebFetch, NotebookEdit, Task) | 4 (read, write, edit, bash) + 3 optional (grep, find, ls) | Both |
| Agent Architecture | Monorepo TypeScript CLI — single package with built-in tool execution, sub-agents, and team coordination | 4-package monorepo (pi-ai, pi-agent-core, pi-tui, pi-coding-agent) — modular separation of LLM abstraction, agent loop, TUI, and CLI | Both |
| Sub-Agent Support | Native Task tool — 7 parallel sub-agents, permission inheritance, typed agent roles (Explore, Plan, Bash, general-purpose) | None built-in, but available through extension that spawns separate pi processes in single/parallel/chain modes with different models per sub-agent | Claude Code |
| Agent Teams | Native team coordination (lead + workers, shared task lists, message passing, broadcast) | None built-in, but achievable through SDK orchestration scripts or RPC mode driving multiple pi processes | Claude Code |
| Default Permission Model | 5 modes (default, plan, acceptEdits, bypassPermissions, dontAsk) — deny-first with filesystem/network sandbox | None by default ("YOLO mode") — runs everything without asking. Permission-gate extension available but opt-in | Claude Code |
| Memory File | CLAUDE.md (project root, nested dirs, user-level) — auto-loaded, hierarchical | AGENTS.md — similar convention, compatible with ~/.claude/skills cross-tool standard | Tie |
| Cost Visibility | Available via /cost command, customizable via statusline configuration | Immediately visible in footer by default, further customizable via extensions and getSessionStats() API | Tie |
| Hooks System | Shell-command hooks (PreToolUse, PostToolUse, Stop, Notification) — external scripts, pass/fail | TypeScript extension events (20+ types) — in-process async handlers that block, modify, transform, access session state, render UI | Pi |
| Session Format | Linear conversation | JSONL tree with id/parentId (branching, forking, labels via /tree and /fork) | Pi |
| Extension State | No built-in state persistence for extensions | pi.appendEntry() persists custom data to session, survives restart | Pi |
Tools & Capabilities
Built-in Tools (Tool-by-Tool)
| Tool | Claude Code | Pi Agent | Winner |
|---|---|---|---|
| Read | Built-in — reads files with optional offset/limit, images, PDFs, notebooks | Built-in (read) — reads files with optional range, auto-resizes images |
Tie |
| Write | Built-in — creates or overwrites files | Built-in (write) — creates or overwrites files |
Tie |
| Edit | Built-in — exact string replacement with replace_all option | Built-in (edit) — surgical find-and-replace, returns unified diff |
Tie |
| Bash | Built-in — shell execution with timeout, background mode, description | Built-in (bash) — shell execution with streaming output and abort |
Tie |
| Glob | Built-in — fast file pattern matching, sorted by modification time | Not built-in. Optional find tool available via --tools flag |
Claude Code |
| Grep | Built-in — ripgrep-powered search with regex, context lines, output modes | Not built-in by default. Optional grep tool available via --tools flag |
Tie |
| WebSearch | Built-in — web search with domain filtering, returns formatted results | Not built-in, customizable via extensions | Claude Code |
| WebFetch | Built-in — fetches URL content, converts HTML to markdown, AI processing | Not built-in, customizable via extensions | Claude Code |
| NotebookEdit | Built-in — Jupyter notebook cell editing (replace, insert, delete) | Not built-in, customizable via extensions | Claude Code |
| Task (Sub-agents) | Built-in — spawns typed sub-agents (Explore, Plan, Bash, general-purpose) with parallel execution | Not built-in, customizable via extensions. Subagent extension spawns separate pi processes | Claude Code |
| ls | Not a dedicated tool (use Bash or Glob) | Optional built-in (ls) via --tools flag |
Tie |
Tool System Capabilities
| Feature | Claude Code | Pi Agent | Winner |
|---|---|---|---|
| Tool Observability | Sub-agent tool calls opaque | Every tool call, token, and dollar visible | Pi |
| Custom Tools | Via MCP servers (external process, JSON-RPC) | pi.registerTool() in-process TypeScript, streaming results, custom rendering | Pi |
| Tool Override | Not possible | Register tool with same name to replace built-in (e.g., audited read) | Pi |
| MCP Support | Native first-class, lazy loading (95% context reduction), OAuth | Not built-in (by design, argues 7-14k token overhead); available via extensions | Claude Code |
| Tool Count Philosophy | More tools = more capable out of the box (10+) | Fewer tools = smaller system prompt (~1000 tokens), trusts frontier models | Both |
Hooks & Event System
Claude Code: 14 hook events, 3 handler types (command, prompt, agent) — shell-based, JSON stdin/stdout Pi: 25 extension events across 7 categories — in-process TypeScript with full API access
Architecture
| Feature | Claude Code | Pi Agent | Winner |
|---|---|---|---|
| Hook Language | Shell commands (any language), LLM prompts, or agent subprocesses | TypeScript (in-process, zero-build via jiti) | Pi |
| Handler Types | 3: command (shell), prompt (LLM eval), agent (multi-turn subagent) | 1: async TypeScript handler with full session/UI access | Both |
| Hook Configuration | JSON in settings files (.claude/settings.json, managed policy, plugin hooks.json, skill/agent frontmatter) | TypeScript code in extension files | Both |
| Can Modify Tool Input | Yes — updatedInput in PreToolUse/PermissionRequest | Yes — return modified args from tool_call handler | Tie |
| Async/Background Hooks | Yes — async: true on command hooks (non-blocking, results delivered next turn) | Yes — handlers are async by default, can fire-and-forget | Tie |
| Inter-Hook Communication | No built-in | Yes — pi.events shared event bus between extensions | Pi |
Hook-by-Hook Mapping
| Lifecycle Point | Claude Code Hook | Pi Extension Event(s) | Notes |
|---|---|---|---|
| Session starts | SessionStart (matcher: startup/resume/clear/compact) |
session_start |
CC can persist env vars via CLAUDE_ENV_FILE. Both can inject context |
| User submits prompt | UserPromptSubmit — can block prompt, add context |
input — can block, transform text, or handle entirely |
Pi also distinguishes source: interactive/rpc/extension |
| Before tool executes | PreToolUse — allow/deny/ask, modify input |
tool_call — block with reason, modify args, typed per-tool |
Both can intercept and modify. Pi has typed narrowing via isToolCallEventType |
| Permission dialog shown | PermissionRequest — auto-allow/deny on behalf of user |
N/A (Pi has no permission system by default) | CC-only — Pi runs YOLO by default, permission-gate is an extension |
| After tool succeeds | PostToolUse — feedback to Claude, modify MCP output |
tool_result — modify results, log, transform output |
Comparable |
| After tool fails | PostToolUseFailure — add context about the failure |
tool_result (isError flag) |
CC has a dedicated event; Pi uses same event with error flag |
| Tool execution streaming | N/A | tool_execution_start, tool_execution_update, tool_execution_end |
Pi-only — real-time streaming of tool execution progress |
| Bash spawn intercept | N/A | BashSpawnHook — modify command, cwd, env before bash executes | Pi-only — intercepts at process spawn level |
| User runs bash directly | N/A | user_bash — fired when user types shell commands (!! prefix) |
Pi-only |
| Notification sent | Notification (matcher: permission_prompt/idle_prompt/auth_success/elicitation_dialog) |
N/A (use ctx.ui.notify in any handler) | CC-only as a hook event |
| Subagent spawned | SubagentStart (matcher: agent type) |
N/A (Pi has no built-in subagents) | CC-only |
| Subagent finished | SubagentStop — can prevent subagent from stopping |
N/A | CC-only |
| Agent stops responding | Stop — can force Claude to continue |
N/A (use turn_end or agent_end to react) | CC can block stopping; Pi can't but can queue follow-ups |
| Teammate goes idle | TeammateIdle — can force teammate to continue |
N/A (Pi has no built-in teams) | CC-only |
| Task marked complete | TaskCompleted — can block completion |
N/A (Pi has no built-in task system) | CC-only |
| Before compaction | PreCompact (matcher: manual/auto) |
session_before_compact — can provide custom compaction entirely |
Pi can fully replace compaction logic |
| After compaction | N/A | session_compact |
Pi-only |
| Before session branching | N/A | session_before_fork, session_before_switch, session_before_tree |
Pi-only — session tree architecture |
| After session branching | N/A | session_fork, session_switch, session_tree |
Pi-only |
| Session ends | SessionEnd (matcher: clear/logout/exit/other) — no decision control |
session_shutdown |
Both fire on exit; neither can block |
| Before agent processes prompt | N/A | before_agent_start — can modify system prompt, images, prompt text |
Pi-only — dynamic system prompt per-turn |
| Agent turn lifecycle | N/A | agent_start, agent_end, turn_start, turn_end |
Pi-only — granular agent lifecycle |
| Message streaming | N/A | message_start, message_update, message_end |
Pi-only — token-by-token streaming access |
| Model changed | N/A | model_select (source: set/cycle/restore) |
Pi-only — react to model switches |
| Context window access | N/A | context — deep copy of messages, can filter/prune |
Pi-only — direct context manipulation |
Extensions & Customization
| Feature | Claude Code | Pi Agent | Winner |
|---|---|---|---|
| Extension Language | Shell scripts (hooks), Markdown (commands) | TypeScript (zero-build via jiti) | Pi |
| Slash Commands | .claude/commands/*.md prompt templates | Prompt templates + /skill:name + extension-registered commands | Tie |
| Package Distribution | Plugin marketplace — /plugin commands, git-based sharing |
pi install npm:/git:/local, pi config TUI, npm gallery | Both |
| Skills (Agent Skills Standard) | Yes (auto-invocation) | Yes (progressive disclosure, cross-tool compat with ~/.claude/skills) | Tie |
| Themes | Minimally customizable | 51 color tokens, hot-reload, dark/light built-in, community themes via packages | Pi |
| Custom Keyboard Shortcuts | ~/.claude/keybindings.json | pi.registerShortcut() in extensions | Tie |
| Custom CLI Flags | Not possible | pi.registerFlag() adds custom flags to CLI | Pi |
| Custom Providers | Not possible | pi.registerProvider() with OAuth support | Pi |
| Custom Editors | Not possible | Modal editor (vim), emacs bindings, rainbow editor via extensions | Pi |
UI & Terminal
| Feature | Claude Code | Pi Agent | Winner |
|---|---|---|---|
| Custom Header | No | ctx.ui.setHeader() replaces logo/keybinding hints with custom component | Pi |
| Custom Footer | Configurable statusline (tokens, cost, model) | ctx.ui.setFooter() with git branch, token stats, cost tracking, anything | Pi |
| Status Line | Configurable statusline (tokens, cost, model) | ctx.ui.setStatus() with themed colors, turn tracking, custom data | Pi |
| Widgets | No | ctx.ui.setWidget() above/below editor with custom content | Pi |
| Overlays | No | Full overlay applications (Doom, Space Invaders, QA test overlays) | Pi |
| Dialogs | Basic permission prompts | ctx.ui.select(), confirm(), input(), editor() + custom rendering | Pi |
| Rendering | Standard terminal with known issues | Standard terminal with known issues | Both |
| Message Queuing | Supported — queue messages while agent works | Enter = steer (interrupt), Alt+Enter = follow-up (queue after completion) | Both |
Programmatic & SDK
| Feature | Claude Code | Pi Agent | Winner |
|---|---|---|---|
| Non-Interactive Mode | claude --print | pi -p (+ stdin auto-activates) | Tie |
| JSON Streaming | --output-format stream-json | --mode json (JSONL events with full lifecycle) | Tie |
| RPC Mode | None | --mode rpc (26+ commands, bidirectional JSON protocol, any language) | Pi |
| Node.js SDK | @anthropic-ai/claude-agent-sdk | @mariozechner/pi-coding-agent (createAgentSession, full internal API) | Tie |
| Mid-Stream Control | ClaudeSDKClient.interrupt() — stop and redirect | steer() interrupts, followUp() queues messages while agent works | Pi |
| Session Stats API | Limited | getSessionStats() returns tokens (in/out/cache), cost, tool calls per session | Pi |
| HTML Export | /export — session to HTML | --export, /export, session.exportToHtml() | Both |
| SDK Examples | Docs-based | 12 official examples from minimal to full-control in package | Pi |
Multi-Agent & Orchestration
| Feature | Claude Code | Pi Agent | Winner |
|---|---|---|---|
| Sub-Agents | Native Task tool, 7 parallel, permission inheritance | Subagent extension (single/parallel/chain modes), spawns separate pi processes | Claude Code |
| Agent Teams | Native team coordination (lead + workers) | No built-in equivalent; use orchestration scripts | Claude Code |
| Multi-Model Orchestration | Not possible (single provider) | Different models per sub-agent (scout on flash, worker on opus) | Pi |
Enterprise & Platform
| Feature | Claude Code | Pi Agent | Winner |
|---|---|---|---|
| IDE Integration | VS Code, JetBrains, Cursor (inline diffs, @mentions) | Terminal-only (could integrate via RPC) | Claude Code |
| Web/Mobile/Desktop | claude.ai/code, iOS app, desktop app | Terminal only | Claude Code |
| Enterprise SSO/Audit | Yes (SSO, MFA, audit logs, admin dashboard) | No | Claude Code |
| Permissions/Sandboxing | 5 modes, deny-first rules, filesystem/network sandbox | None by default ("YOLO mode"); permission-gate extension available | Claude Code |
| Git Integration | Deep (commits, PRs, merge conflicts, GitHub Actions, GitLab CI) | Via bash; git-checkpoint extension available | Claude Code |
| Slack/Chat Integration | Native @Claude mentions to PRs | pi-mom Slack bot package | Claude Code |
Sharing & Distribution
| Feature | Claude Code | Pi Agent | Winner |
|---|---|---|---|
| Package System | Plugin marketplace — /plugin commands, .claude-plugin/plugin.json manifest |
pi install npm:/git:/local — package.json with pi key, pi-package npm keyword |
Both |
| What's Bundled | Skills, agents, hooks, MCP servers, LSP servers | Extensions, skills, prompt templates, themes | Both |
| Distribution Sources | Marketplace (GitHub repo, git URL, npm, pip, direct URL, local path) | npm registry, git (GitHub/GitLab/SSH), local paths — no intermediate marketplace needed | Both |
| Discovery | Official claude-plugins-official marketplace + team/community marketplaces via extraKnownMarketplaces |
npm search (pi-package keyword) + gallery at shittycodingagent.ai/packages with video/image previews |
Both |
| Scope | User, project, local, managed (enterprise) — namespaced as plugin-name:skill-name |
Global (~/.pi/) or project (.pi/, -l flag) — project settings auto-install missing packages on startup |
Tie |
| Config UI | /plugin slash commands for install/browse/manage |
pi config interactive TUI for enable/disable per-resource |
Tie |
| Try Without Installing | No equivalent | pi -e npm:@foo/bar — ephemeral install for current session only |
Pi |
| Cross-Tool Portability | Agent Skills standard (agentskills.io) — shared with VS Code, Codex, Cursor, GitHub | No cross-tool standard — Pi-specific extensions | Claude Code |
| Enterprise Controls | strictKnownMarketplaces, allowlists by repo/URL/host regex, managed plugin deployment |
No enterprise controls — trust-based, review source before installing | Claude Code |
| Git-Based Sharing | .claude/ directory (settings, skills, agents, rules, hooks) committed to repo — team gets config on clone |
.pi/settings.json with packages — team gets packages auto-installed on startup |
Tie |
| Update Mechanism | Marketplace auto-updates at startup (configurable) | pi update for non-pinned packages, version pinning with @version |
Tie |
| Package Filtering | Plugin resources loaded as-is (namespaced to prevent conflicts) | Glob patterns + !exclusions per resource type, force-include/exclude exact paths |
Pi |
Community & Ecosystem
| Feature | Claude Code | Pi Agent |
|---|---|---|
| Creator | Anthropic — $1B+ ARR, enterprise AI company | Mario Zechner — libGDX creator (24.8K stars), solo maintainer |
| Traction | Enterprise adoption, deep IDE integrations, massive user base | 11.5K stars, 3.17M monthly npm downloads, 208 versions |
| Endorsements | Enterprise customers, Anthropic ecosystem | Armin Ronacher (Flask/Ruff) uses + contributes, powers OpenClaw (145K stars) |
| Release Velocity | Regular releases | 10+ releases in 8 days, new model support within hours |