The AI runs on us.
Every other agent bills you for the model — a subscription, your own API keys, or a meter ticking on every token. Picasso doesn't. Sponsors fund the inference, so you get frontier models inside a beautiful editor for exactly zero.
Bring nothing.
No API key, no subscription, no card. Install in one command and the agent is already wired to real models.
Sponsors cover the tokens.
Inference is paid for by clearly disclosed sponsors — honest, labeled placements, never your wallet. That's what funds the free tier.
Walled off from your work.
Sponsors never see your code and never influence a single token of model output. Your code is never training data or sponsor inventory, and session telemetry and feedback uploads are off by default.
Prefer your own? Drop in an API key or run local models anytime — the free path is a choice, not a cage.
Save routes through
approvals. Always.
App.swift — explain, refactor, or generate a patch. Picasso stays read-only until you approve a write.App entry point. Lines 5–9 build the window group; the run store is injected as an environment object.The agent does not
start cold.
Picasso retrieves relevant memory before planning, injects it into context before it acts, and writes back what mattered after the run — facts, style, decisions, failures.
Give long work
a finish line.
Hand Picasso an objective, completion criteria, permissions, and a budget. It plans, acts, observes, reflects, repairs failures, and keeps going until the work is done or a boundary is reached.
Plan. Act. Observe. Reflect.
The agent loop is visible by design — across a full-black stream-first CLI and a three-zone macOS workspace with 1M-token context windows, auto-diagnostics on every edit, and isolated git worktrees per thread.
Read-only plan mode
Architecture, entry points, data flow, risks, and verification — grounded in your real workspace, with writes blocked until you approve execution.
Auto-diagnostics in the edit loop
Syntax and typecheck feedback attaches directly to file edits in TypeScript, JavaScript, Python, JSON, and Swift so the agent repairs issues in the same turn.
Git worktree isolation per thread
Use /worktree in the CLI or the Worktree pill in the Mac composer so parallel threads never step on each other's files and merge cleanly back to your branch.
Three-zone macOS workspace & stream-first CLI
Multi-thread sidebar with unread completion indicators and native notifications, instant runs without project setup friction, inline syntax-highlighted diffs, a full-black stream-first CLI with Ctrl+P to jump to any turn, and an inspector with Reasoning, Changes and Agents tabs.
Local branch & diff review
Run /diff main and /review-branch to inspect branch changes, surface regressions, and review syntax-highlighted diffs before opening a pull request.
Agent teams that ask like you do
Named and forked workers, agent teams with a shared task board, MCP client/server tools and hooks. Delegated work follows your approval settings, and /model and /provider switch in one step.
Every agent in view.
The team, its task board and its shells sit beside the thread. When Auto review sends a call back, the approval card says why.
Recently shipped.
What changed in the Mac app and the terminal.
- Mac
Watch the team work.
When Picasso runs a team of agents, the inspector shows the roster and a task board layered by dependency, with each task's owner and whether it is ready or waiting on others.
Learn more - Mac
See the shells the agent opened.
Terminals the agent starts are listed read-only in the inspector with their state and folder, so no shell runs out of sight.
Learn more - Mac
Auto mode, with a reason.
In Auto, a second model reviews routine actions. When it sends one back to you, the approval card says why.
Learn more - CLI + Mac
Subagents ask like you do.
Work delegated to a subagent now follows your approval settings instead of being refused behind the scenes.
Learn more - Mac
Open the document, not the diff.
PDFs, images and Word, Excel and PowerPoint files preview inside the thread. Office files are converted to PDF for viewing.
Learn more - Mac
Pick a folder on another machine.
Attached to a remote engine, you choose the working folder on that machine, and create one if you need it, instead of typing a path.
Learn more - Mac
Questions that keep time.
When the agent asks and will carry on without you, the card counts down and pauses while you write your answer.
Learn more - Mac
- TUI
Jump to any turn.
Press Ctrl+P, type to filter, and land on the turn you want in a long session.
Learn more - TUI
A fuller activity line.
Permission changes, folder changes, hook results, goal updates and preset changes show up as quiet notes in the stream, and parallel code cells show live while they run.
Learn more - Mac
Personas that shape the reply.
Concise, Mentor and Reviewer now change how Picasso answers.
Learn more - CLI + Mac
Nothing phones home by default.
The engine's own session telemetry and feedback uploads are off on both surfaces. Tests capture every outbound connection during a turn and a feedback action to prove it.
Learn more
On launch, the Mac app checks its bundled engine against the manifest it shipped with and won't start if anything has changed.
The important state
stays visible.
1M context windows, edit-loop auto-diagnostics, git worktree isolation, local branch review, runtime memory, and autonomous goals — inspectable in the CLI and the three-zone Mac app.
| Claude Code | Codex CLI | Cursor | Devin | Picasso | |
|---|---|---|---|---|---|
| Context window | Anthropic | OpenAI | Model-dependent | Hosted | Up to 1M ctx |
| Read-only plan mode | Yes | Yes | Yes | Yes | Yes |
| Auto-diagnostics on edits | Partial | No | Yes | Hosted | TS · JS · Py · JSON · Swift |
| Git worktree isolation | Flag / manual | Partial | Partial | Cloud VM | /worktree + Mac Worktree pill |
| Local branch review | Prompted | Yes | Partial | PR-based | /diff main · /review-branch |
| Subagents | Yes | Limited | Yes | Yes | Named + fork |
| 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 |
| Free path | Trial | Limited | Trial | Credits | Sponsor-supported |
| Tool subscription | Claude plans | ChatGPT plans | $20/mo | From $20/mo | $0 |
Four ways in. Pick yours.
Managed inference, a supported provider subscription, your own API keys, or local models through Ollama and vLLM. macOS, Linux, and Windows.
One binary.
No subscription required.
Zero account to start — bring your own keys, or unlock managed inference through honestly disclosed sponsors.
Install opens the day the doors do.
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. Be first on the canvas.