Proto desktop agent console

Proto is a local-first Electron desktop app and CLI agent harness. Open a workspace, select a model, choose a permission mode, and give it a task. The runtime can call actions, read their results, and continue for up to 40 iterations.

Runtime

Read-only actions may run in parallel; write actions run serially.
Cancellation stops the current run cleanly.
Sub-agents have separate context and a smaller iteration cap.
Background tasks and fan-out jobs handle bounded or repeated work.
Context compression keeps long sessions within the selected model’s limits.

Local data

Workspace instructions: PROTO.md, AGENTS.md, or CLAUDE.md
Direct memory file: .proto/MEMORY.md
Indexed memory: .proto/memory/
Session records: .proto/sessions/*.jsonl
Global Plugin config: ~/.proto/mcp.json
Workspace Plugin config: <workspace>/.proto/mcp.json

Files, credentials, memory, conversation history, and session records stay local. Hosted-model requests send the messages and selected context required for that request to the chosen provider. Plugin actions send requested data to the configured Plugin endpoint.

Action families

Files: search, read, write, patch, multi-file edit, and undo
Development: shell, Git, GitHub, diagnostics, definitions, references, and hover
Web: search, page reading, in-app browser, and opt-in computer control
Work management: plans, tasks, goals, sub-agents, background tasks, and fan-out
Content: documents, slides, images, messages, and meetings
ERP.AI: apps, tables, columns, records, views, roles, workflows, reports, pages, documents, branches, previews, API keys, hosting, auth, and portals
Plugins: deferred external actions discovered when needed

Built-in skills

The current desktop build includes 22 skill manifests:

build-app, build-document, build-host, build-page, build-slides
build-workflow, test-workflow
coding, commit, review, pr-review, qa-page-ui
computer-use, deep-research, generate-image, fan-out
erpai-api, erpai-external-apps, erpai-records, import-data
manage-filters, manage-roles

Install additional skills from a registry source, a GitHub repository, a URL, or a local folder.

Permission modes

Mode	Behaviour
Ask	Pause before every action
Auto accept edits	Allow routine file edits; ask for higher-risk actions
Plan	Remove write actions from the action registry
Bypass	Skip approval prompts; use only in an isolated environment

Dangerous command patterns still trigger checks. Root-level destructive patterns such as recursive deletion of / are hard-blocked.

Plugins

Plugins use Model Context Protocol internally. Proto supports local stdio and remote HTTP/SSE connections, OAuth, and environment-variable expansion. Action schemas are deferred and loaded through action search only when the task needs them.

Models

Hosted routing: OpenRouter and Fireworks with your own key
Local providers: Ollama and LM Studio
Model families depend on provider availability and include Claude, Gemini, GLM, Minimax, Qwen, and Llama

Install

Proto is proprietary and in v0 alpha. It is free during alpha; hosted-model usage is billed by the provider attached by the user.