This is the complete historical feature baseline. The following describes capabilities already present in this version; later entries record subsequent changes. Actual execution depends on configured models, installed tools, permissions and an online runtime.
Conversations, long-horizon work and collaboration
- Desktop workspace, project conversations and streamed model interactions display user messages, agent replies, tool calls and results.
- Missions divide larger goals into Tasks with explicit responsibility, retaining requirements, decisions, interactions, context, dependencies and delivery records. Each Task uses a fixed squad and declared workflow.
- Independent work proceeds in parallel; dependent work exchanges accepted typed artifacts. Planning, implementation, research, testing and review can form one delivery chain.
- Tasks support follow-ups, cancellation, retries, replanning and recovery. Recurring work retains history; background execution requires an online local or managed service.
Expert Squads, methods and evolution
- Squads package professional roles, instructions, Skills, tools, MCP servers and workflows as self-contained capabilities selected/configured/installed for projects.
- Existing squads cover coding, frontend replication/innovation, review/debugging, deep research, stock research, squad authoring and Evolution Lab. Research roles cover planning, sources, writing, knowledge organization and citation review; stock roles cover facts, fundamentals, valuation, theses and reports.
- Squad SDK supports creating, validating, versioning and installing custom squads. Algorithms/external agents participate through explicit adapters, contracts, evidence and artifacts.
- Evolution Lab provides candidate revisions, trials, observation, evaluation and feedback. Users review changes, and experiment results/sources remain for comparison.
- Skills are used by role/Task and integrate with existing external-host, Hermes and OpenClaw Skill entries.
Models, tools and execution
- Multiple Providers/models, model selection, service settings and authentication are supported. Compatible Provider plugins participate in streamed tool calls.
- File read/write/edit, patches, directory/code search, commands, planning, todos, questions, waiting and memory are subject to actual role grants and permission boundaries.
- Experimental LSP tools query definitions, references, symbols and call hierarchies, requiring the then-current experiment flag, grants and language servers. Later releases retired this path.
- Browser capabilities cover page actions, web search/fetching, screenshots and observable browser evidence. Computer Use connects host applications for authorized desktop actions.
- Local code/repository work manages directories, Git and isolated worktrees. External coding executors integrate supported tools such as Codex and Claude Code.
- MCP supports external tools/interactive apps, and Agent Client Protocol/plugin entries connect other agent capabilities. Unconfigured executors/integrations require their actual environment prerequisites.
Artifacts, editing and reviewable delivery
- Typed, content-addressed artifacts retain exact lineage, sources and acceptance evidence. Research, planning, development and independent review can cite the same verifiable delivery.
- Files/directories serve as request attachments and artifact references; conversations, Missions, Tasks and right panels show the corresponding context and outputs.
- The workspace includes a source editor and terminal, with Markdown, HTML, Portable Document Format (PDF), code, charts, maps, relationship graphs and interactive-app readers.
- OfficeCLI and editable document/spreadsheet/presentation artifact processing are available. Sheets, slides, charts and graphics can be viewed/operated in their components.
- Independent acceptance checks implementation, runtime, visuals, sources and artifacts. Missing evidence, required permissions and actual blockers retain explicit states; delivery records trace stage contributions.
Persistence, memory, permissions and observation
- Tasks, requirements, dispatch lineage, interactions, artifacts, acceptance evidence, Session state and project knowledge persist in local SQLite for reading/recovery after restart.
- Project/Session and global memory/preferences inform later work. Compaction and Session records retain information required for long-horizon work.
- Questions and Permissions support human responses; tools/external operations follow approved scope. Operator feedback can enter subsequent Task work.
- Model calls, tool activity, tokens and cost are recorded alongside logs, diagnostics and model/MCP checks to investigate actual execution.
Desktop, services, channels and developer integration
- Tauri desktop shell and local web workspace are available. The repository provides desktop/CLI packaging for Windows x64, macOS x64/ARM64 and Linux x64/ARM64, with platform-format guidance.
- Headless HTTP APIs manage Task lifecycle and provide progress through Server-Sent Events (SSE).
- CLI supports services, models/authentication, MCP, Session import/export, statistics, configuration checks and doctor diagnostics. JavaScript SDK and GitHub Actions support development/repository automation.
- Slack gateways start work from threads, synchronize progress, process allow/always/reject permission responses and send follow-ups. Adapters for Telegram, Discord, Feishu, WhatsApp, Google Chat, Microsoft Teams, Line, Matrix, Mattermost, Signal, WeCom and DingTalk require service configuration before use.
- English/Chinese guides cover development, builds, packaging, configuration and integrations. The project is open source under the MIT license.
Version boundary
- These notes describe the 0.0.35beta source capability baseline. Historical installer availability follows its release page. Signed desktop updates, hosted markets, global Chat, side conversations and references are recorded in the later versions that actually introduced them.