ClawAI Coding Agent 1.96.0

1.96.0

The agent can now do much more of a real engineering job on its own. Every new tool below is off by default and needs an explicit flag, so a run that does not ask for them sees exactly the tools it saw before. They work in the command line (clawai) and the host-free SDK; the editor panel is unchanged except where noted.

• Browser (--allow-tools browser): the agent can open a page in a hidden browser, read it, click, type, take a screenshot and read console and network errors, to check the app it just built. Private and local addresses stay blocked unless you name them with --browser-allow-host.
• HTTP requests (--http-allow-host, --allow-tools http,http-write): the agent can call your project's own API to test it. It can only reach hosts you list, and a request that changes data (POST, PUT, PATCH, DELETE) is asked about first in every permission mode.
• Task plan (--task-plan, --plan-file, --require-plan): the agent keeps a step plan, and a step with a check can only be marked done when its check passes, so it cannot declare success early.
• Background processes (process.watch, part of the command grant): start a dev server or a watch-mode test, wait for a line of output, read what it printed since, and stop it. Everything is stopped when the run ends.
• Gate runner (code.gates, part of the command grant): find and run the project's lint, typecheck, test and build, and get a short structured result instead of a page of logs. --done-check-gates lint,typecheck,test makes those gates the done check.
• Shell (--allow-tools shell and --allow-shell, both needed): a real shell for pipes and &&. Every script is shown to you for approval first, and a safety screen refuses the dangerous ones.
• Repository knowledge (--load-knowledge): the agent reads your project's instruction files (such as CLAUDE.md, AGENTS.md, rules and skills) on a new task and can search them.
• Pictures (--image, --vision): the agent can look at a screenshot or an image in your workspace.
• Helper agents (--allow-tools agents, --max-agents): the agent can start sub-agents on parts of a big job. A helper can never hold more permissions than the agent that started it.
• Runner and panel: a routine can hand the runner secrets that only commands see and never the model, the runner reports its version and platform, and attaching to a runner session continues where it left off. A message sent while a run waits for your approval now says so, with a button to the approval, and the secondary side-bar sections start collapsed so the chat keeps the space.
• Safety: choosing a permission mode never switches a new tool on by itself, and plan mode removes all of them.

Verified release

• All split CI gates passed for commit .
• The attached VSIX is the exact final artifact produced after every CI gate passed.
• The same VSIX is published to the Visual Studio Marketplace.

Installation

Download , run Extensions: Install from VSIX... in VS Code, select the file, and reload the window.