The most beautiful coding agent ever made, with a real editor inside. It reads a repo, maps the architecture, plans without touching files, edits when approved, runs commands, writes tests, debugs failures, and remembers what matters.
The CLI, TUI, and Mac app share the same runtime state: memory, plans, goals, approvals, replay, subagents, automations, voice, provider routing, and permission modes. And it's free for developers — that's the point.
Plan mode is read-onlyMemory is runtime context/goal proves it's doneReplay shows the proof
Picasso — Studio
$0
for individual developers on managed access.
4
credential modes. Managed, subscription, API key, and local.
100%
of writes route through approvals by default. Loosening that is your call, never the agent's.
/goal
autonomous work with criteria, budget, approvals, and replay evidence.
Picasso for macOS
A desktop app you can actually see.
The same agent, the same boundaries — now with a native Mac surface. Compose runs, speak the prompt, watch the agent work step by step, and let long-running goals drive themselves to a verified finish. In the default mode, nothing touches your files without passing through the approval queue.
Picasso — Goals
Goals that finish themselves.
Hand Picasso a goal and it drives the loop — iterating, checking deterministic completion criteria, and running an independent audit before it calls anything done. The CLI and the Mac app run the exact same engine, so a goal you start in the terminal continues on your desktop.
Picasso — Plan
Plan before it touches anything.
Plan mode reads your repo and lays out the change as a roadmap — architecture, impact, and the exact moves — grounded in your real workspace, without writing a single file until you say so.
The trust loop
Save routes through approvals. Always.
AuthMiddleware.swifteditor
⌘SSave AuthMiddleware.swift
queuedwrite held — nothing touched disk
01You save in the editor — the write is queued, not applied.
approval queue1 pending
- let hash = bcrypt.hash(password)+ let hash = argon2id.hash(password)
RejectRevise✓ Approve
02It lands in the approval queue with the exact diff to review.
AuthMiddleware.swiftrecorded
✓ approvedby you · written to disk
+4−1 · 1 file · replay available
03Approve, and only then is it written to disk — and recorded.
Picasso retrieves relevant memory before planning, injects it into model context before the agent acts, and writes useful facts, style, decisions, failures, workflow notes, repo rules, and model-performance observations after the run.
The structured store is the source of truth. Human-readable memory files mirror it so users can inspect what Picasso remembers without turning Markdown into a fragile database.
Runtime wired.Normal CLI, TUI, and Mac runs retrieve and write memory.
Retrieval only.Your code is not training data or sponsor inventory.
User-visible.Inspect, edit, pin, export, delete, or clear records.
Failure aware.Repeated mistakes can become memories the next run avoids.
Some work takes more than one pass. Give Picasso an objective, completion criteria, permissions, and a budget. It plans, acts, observes, reflects, repairs failures, and keeps going until the work is complete or a boundary is reached.
Goals carry provider calls, tool calls, tests, cache events, memory writes, approvals, pauses, resumes, cancellation, and completion evidence across CLI, TUI, and Mac.
The agent loop is visible by design. You see the plan, the tools, the failures, the checks, and the reason the work changes direction.
01
Read-only plan mode
Architecture, entry points, dependencies, data flow, risks, files, and verification, with writes blocked until you approve execution.
02
Subagents and replay
Named and forked workers can investigate in parallel, then replay shows the plans, tools, approvals, edits, and results.
03
Modes, MCP, and hooks
Permission mode, MCP tools, hooks, sponsor boundaries, and provider routing stay visible instead of hiding inside a prompt.
Visibility
The important state stays visible.
Plans, canvases, goals, memory, subagents, permissions, budgets, provider routes, command evidence, and replay are inspectable in the terminal, the TUI, and the Mac app. Sponsor disclosures stay visible and separate from the agent loop.
Capability
Claude Code
Codex CLI
Cursor
Devin
Picasso
Read-only plan mode
Yes
Yes
Yes
Yes
Yes
Subagents
Yes
Limited
Yes
Yes
Named + fork
MCP
Yes
Yes
Yes
Yes
Client + server
Multi-provider
Yes
No
Yes
No
Yes
Runtime memory
Project files
No
Partial
Wiki
Retrieve, inject, write back
Autonomous /goal
Anthropic
OpenAI
No
Yes
All routes
Replay evidence
Partial
Partial
No
Hosted
CLI, TUI, Mac
OS sandbox
Workspace
Workspace
No
Cloud
Yes
Free path
Trial
Limited
Trial
Credits
Sponsor-supported
Tool subscription
Claude plans
ChatGPT plans
$20/mo
From $20/mo
$0
◇ Sponsored·Builder patron·Sponsor-supported managed access.·Learn why this appears.Sponsors never affect code, plans, tool use, or model output.
Install
Four ways in. Pick yours.
Use managed inference, a supported provider subscription, your own API keys, or local models through Ollama and vLLM.
macOS, Linux, and Windows. No Picasso subscription required for individual developers.
$npm i -g @picasso/climanaged path
brewbrew install picassocopy
npmnpm i -g @picasso/clicopy
pnpmpnpm add -g @picasso/clicopy
windowswinget install picasso.picassocopy
The waitlist
Code Freely.Code Beautifully.
Picasso for Mac is almost here — a coding agent that looks the way serious tools should, and costs what creative freedom should: nothing. Leave your email and be first on the canvas.
Sponsors and labs — the early canvas is yours. Choose Sponsor or Lab above and we'll reach out before launch.
