Features

Enter /side your question, or select text and choose “Ask in side chat,” to start an independent chat from the main conversation’s completed history. Multiple side chats can be collapsed, restored, and run alongside the main chat; Plan and Workflow modes are not yet supported in side chats.

Right-side panels organize files, Workspace, diffs, plans, Canvas, and the browser into tabs that can be reordered, maximized, or collapsed. Projects can link multiple source folders. File breadcrumbs navigate directories, and HTML previews switch between source and a rendered view without scripts. Selected text can be copied or quoted into the composer for you to send.

Click a modified-file label below a reply or a workspace filename with a change snapshot to see the saved before-and-after diff. Repeated edits to the same path show the latest change, and new files show all additions. Modified text files without snapshots show “No diff data”; partial snapshots indicate truncation. To see the current contents, choose Preview from the context or ⋯ menu. Media outputs without snapshots and read-only files keep their usual open action.

Cross-chat delivery receipts and source links navigate in both directions, restoring the main chat and side panel when the source is a side chat. Failed context compaction explains the problem and offers a retry. Ordinary long tool output is not guaranteed to remain available in full after compaction; use the tool’s own continuation cursor or rerun it.

Type @ in the composer to reference an existing regular conversation in a new task. When asked, Hope Agent can read its context or send work to that conversation. The session sidebar can also import local Codex history from the machine running Hope Agent. Reimports update records from the same source; imported chats can be searched, viewed, and exported, but cannot be continued.

Hope Agent separates long-running work into five composable roles: Goal defines the outcome and completion criteria, Workflow performs one concrete execution, Loop decides when to continue, Task exposes live progress, and Mode controls execution autonomy.

Goals support budgets, pause, and resume, and completion requires a conservative audit that surfaces result evidence. Workflows dynamically organize phases, conditions, parallelism, multiple agents, tools, diffs, reviews, and verification, with durable run history and conservative crash recovery. Loops advance again on intervals, conditions, internal events, or model-selected wakeups, with budgets, backoff, and no-progress protection.

Complex tasks can begin in Plan Mode with an editable implementation plan, while Tasks track the current step. Long-running tools and sub-agents work in the background and return milestones progressively without blocking the conversation.

Design Space generates websites, mobile prototypes, presentations, dashboards, posters, documents, email, images, motion, audio, and interactive components. Generation streams into a live preview, with project-level AI chat, element editing, annotations, undo/redo, device previews, version history, and an artifact library.

Artifacts export as HTML / PNG / PDF / PPTX / MP4 / ZIP. You can also extract a brand system from screenshots, URLs, Figma, or an existing repository and reuse design tokens across artifacts. Bind a real repository to send an artifact into the main conversation for implementation and surface later code changes back into the design.

Knowledge Space lets you and the AI work on real Markdown files with hybrid full-text and vector retrieval, backlinks, graph view, atomic notes, and reviewable AI organization proposals. Existing Obsidian vaults can be attached, with outside changes synced live.

Long-term memory is organized across Global / Project / Agent scopes. A compact Core stays stable in context while details return on demand through full-text and vector retrieval. Idle-time consolidation can produce a Dream Diary and distill reviewable communication preferences and work habits from history.

Completed complex work can become a draft skill for your review and later reuse. Skills support conditional activation, sub-agent execution, tool allowlists, and the agentskills.io standard. Incognito sessions disable long-term memory, cross-session awareness, and persistence paths, then remove session data when the chat ends.

External vaults are read-only by default. To write, first allow external writes in the space settings, then explicitly mount it as read/write in the chat or project; the setting alone does not grant mount access. Background organization never writes to external vaults. The built-in skill installer prepares a fixed snapshot from a GitHub subfolder or local directory, previews it before installation, and checks the result and dependencies.

On macOS, granted permissions let Hope Agent observe and operate the desktop, windows, menus, keyboard, and pointer. The controllable browser includes a live mirror so you can see the pages it visits and manipulates. Side effects share one approval flow, while login, 2FA, and CAPTCHA steps pause for you.

The built-in MCP client covers major transports and OAuth 2.1. Hooks attach command / HTTP / MCP / prompt / agent handlers to lifecycle events. Deep Feishu/Lark integration adds tools across documents, bitable, drive, wiki, approvals, calendar, contacts, and recruiting.

Preset teams or dynamic sub-agents work in parallel and summarize results back to the main conversation. Natural-language schedules can run recurring work and deliver results to connected IM conversations.

Web reading supports stable snapshots with continuation cursors, dynamic pages, PDFs, and text formats, with source, freshness, and content-budget controls. Settings offers read-only browser, Docker, and toolchain diagnostics. Project Hooks require workspace-specific authorization, and custom sandbox images must use immutable digests.

Hope Agent connects to popular IMs including Telegram, Discord, Slack, and Feishu/Lark. Images, voice, and files enter multimodal context directly, approvals can happen inside the chat app, and sessions hand off between desktop, browser, and IM.

Slack and Telegram now use platform-native streaming: generation continuously updates one reply. Slack can show plans, task progress, and interactive buttons, while Telegram supports native draft streaming followed by rich text, media, and button finalization. Unsupported cases fall back safely, and uncertain deliveries are not duplicated.

If the bound session is still answering the previous message, new messages enter a durable queue and receive replies in send order—even across an application restart.

Configuration, sessions, memory, attachments, skills, and logs live under ~/.hope-agent/ by default, while model requests connect directly to providers. Server mode uses one Owner Token. Same-origin browsers exchange it for an HttpOnly session cookie; the Root Token stays out of URLs and persistent browser storage. Cross-origin clients and preview resources use scoped access, while web fetching checks DNS and SSRF boundaries before redirects.

Sensitive tools enter one approval flow, while high-risk commands and file writes can run inside a Docker sandbox; container deployments support isolated workspaces too. Configuration changes create rollback snapshots, and Guardian, launchd/systemd, and subsystem watchdogs provide restart, diagnosis, and reconnection.

Refreshing or closing a browser no longer interrupts server-owned conversations; reconnecting restores the true run state, with Stop, pending approvals, and background activity sharing one reliability boundary.

Dashboard tracks cost, tokens, activity, health, Plans, and long-running work in one place. Recap reviews a period of conversation history, produces a multi-section report, and exports standalone HTML.

Capability Evaluation runs Goal, Workflow, asynchronous-task, and multi-agent synthetic scenarios with real models, recording completion, tool calls, duration, tokens, and cost with model, version, and trend comparisons. Evaluation runs in an isolated sidecar so it does not interfere with the main application's sessions and tasks.

Before evaluation, set cost, time, token, request, and concurrency limits for the whole experiment and individual scenarios. Invalid or missing evidence does not count as completed work.

A schedule can create a regular chat on every trigger or target the current or another existing chat, following its live context, project, knowledge spaces, and working directory. New-chat tasks can use Project, Fresh Worktree, or Persistent Worktree.

Before saving or running, preflight checks the destination chat, model, project, workspace, permissions, sandbox, and delivery targets. Each run has a navigable chat history, status, errors, and delivery results. Ask the AI why the last run failed, or stop just the current run while keeping future triggers.

Fresh Worktrees can be kept, removed when unchanged, or removed after every run. Worktrees that have been taken over or are still in use are never automatically cleaned up.

The catalog includes 50 provider templates and 477 preset model entries, counted per provider; a model offered by multiple providers appears more than once. It covers Anthropic, OpenAI, Codex, Google, OpenRouter, Requesty, Chinese providers, and local endpoints. Search across providers by model ID to fill capability information, and confirm before applying catalog prices.

Current presets include GPT-6 Astra, Claude Fable 5.1, Gemini 3.8 Flash, and image-capable DeepSeek Flash. Availability depends on the selected provider and account permissions. Adding a connection preserves the existing default model; an unavailable default can fall back to an available alternative. Each provider supports multiple authentication profiles.

For local models, first install the runtime manually from Ollama. Then use Hope Agent Settings to download a model, register the provider, and switch. Models running on your computer do not require a cloud API key.

The Requesty template supports managed routing policies and custom model IDs, with an optional European endpoint. Model availability still depends on the account and gateway.