ClawAI Coding Agent 1.100.0

1.100.0

• Rate limits no longer stall or sink a run. A model that answers 429 is retried 4 times, then the run switches to the next model in clawAI.fallbackModels (CLI: --fallback-model, env CLAW_FALLBACK_MODELS), at the start or in the middle of the run, and the panel says "Switched model". Each model is tried once; with no fallback the run ends with a clear "rate limited" message instead of waiting.

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.

ClawAI Coding Agent 1.99.0

1.99.0

• Safer by design: everything that can reach a website, run a script or start another agent is now asked about in every approval mode. The browser can no longer be tricked into reaching private addresses through WebSockets or look-alike names, a stuck page can no longer freeze the agent, and password fields are hidden from page snapshots.
• Sub-agents and long commands are safer too: a sub-agent cannot spend its parent's budget twice or reach outside its folders, and a script screen, process watcher and code checker were hardened after about 300 attack rounds.
• A plan-file runner: clawai orchestrate --plan plan.json runs several agents in stages, each limited to its own folders and checks, and writes a report.
• Cheaper to run: the tools' descriptions are about a quarter smaller, --tools-profile picks a small set of tools, and --defer-tools loads a tool's full description only when the agent asks for it.
• In VS Code you can now allow the agent to call websites you list and, if you choose, to run scripts you approve one by one (settings clawAI.tools.httpAllowHosts and clawAI.tools.shellEnabled, both off by default).
• A commit that a hook refuses is now reported clearly as not done, and colour codes no longer clutter git output.

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.

ClawAI Coding Agent 1.98.0

1.98.0

• Long agent runs no longer stop after about 15 minutes: the agent renews its sign-in during the run and picks its event stream up where it left off. Callers with their own token can pass a refresh token too.
• A complete, generated list of every tool the agent can use (docs/TOOLS.md), a short FAQ, and step-by-step guides for letting the agent plan, test a page or an API, run long commands and work in parallel. The README has a new "Let the agent test and ship for you" section; everything risky stays off until you turn it on.
• A smaller, faster install: the web-request tool now loads only when it is used, stray files can no longer end up in the package, and the public types for the new options are complete.

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.

ClawAI Coding Agent 1.97.0

1.97.0

• Added the plan for making the agent able to deliver a whole feature on its own (browser testing, API testing, background processes, task plans, sub-agents and more), with what shipped in 1.96.0 and what is still open.

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.

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.

ClawAI Coding Agent 1.95.0

1.95.0

• Creating a folder with the agent now works. Every folder request used to fail with "VS Code rejected the file transaction", even when the folder was created.
• The agent is more forgiving about how a model words a file request: folders can be requested with plain arguments, a new file no longer needs an empty "previous version" field, and a short change label is accepted. Safety checks that protect your existing files are unchanged.
• Asking the agent to look in a folder that does not exist yet now answers "it does not exist yet" instead of showing a technical error with your computer's full path.
• Fewer false warnings: a model that can run tools is no longer marked "cannot call tools" because one list of models forgot to say so, and the Ollama and llama.cpp warnings are hidden when you have not set those up.
• If ClawAI is restarting and the server answers with a "Bad Gateway" page, you now see one plain sentence instead of a block of web page code.

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.

ClawAI Coding Agent 1.93.0

1.93.0

• Maintenance release: the automated checks on GitHub no longer fail when the VS Code test window occasionally does not come up. The extension-host test lane now retries once before reporting a failure. Nothing changes in the extension itself.

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.

ClawAI Coding Agent 1.92.0

1.92.0

Security hardening and a better command line. Commands the agent runs are still not sandboxed: a command can read or write anywhere you can. These changes make the built-in tools and the checks around commands much harder to get around.

• A folder you limit the agent to (write scope) now also covers .git, Windows alternate data streams (file.txt::$DATA) and renames of whole folders, so a limited agent can no longer plant a git hook or edit a protected file through an alias.
• After a command or a git action, the agent now checks for new files in .git (hooks and config) and directly beside your workspace, and removes them. It cannot see a write into a deeper folder or to another drive.
• Windows npm and npx argument quoting is fixed: arguments that could be read as extra commands (", &, |, <, >, ^) are refused.
• Git options that run other programs or write to files outside the workspace are refused.
• Searching with a regular expression or a glob can no longer freeze the agent; patterns that could run forever are refused.
• Approvals are strict: only a real "yes" approves. Anything else counts as "no".
• Numbers given to file tools must be whole numbers.
• Command line: --list-models shows the models you can use; busy servers are retried; the reason a run failed is shown; common tool name variants are understood; and a failing check shows the end of its output.

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.

ClawAI Coding Agent 1.91.0

1.91.0

• Browser sign-in now works when the ClawAI extension runs on a Remote SSH host or another VS Code remote extension host: the local callback is exposed through VS Code's managed forwarding instead of sending the browser to the wrong machine's localhost. The same protection applies to MCP OAuth.
• Every change that reaches main is a release again. The commit hook automatically advances the delivery minor version, keeps the lockfile and README aligned, and CI refuses a main-bound change whose version did not advance.
• CI is split into parallel version, quality, unit, extension-host, browser, and dependency-audit gates. The VSIX and supply-chain files are built only after all of those gates pass.
• A successful main CI run now triggers the Release workflow, which downloads that exact final CI artifact, creates the matching GitHub Release, and publishes the same VSIX to the VS Code Marketplace. Marketplace publication is required rather than silently skipped.

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.

ClawAI Coding Agent 1.90.0

1.90.0

• Compare and Compare + Judge now show their work as it happens. Each model gets its own card the moment it starts, the card shows whether the model is connecting, thinking, writing or finishing, and the answer fills in as it is written. While the judge works you see that it is ranking the answers; when the verdict is ready it appears with the winner, each model's rank and score, the judge's reason for each, and its overall rationale. The cards you were watching stay in place when the run ends.
• Screen readers hear each phase change ("kimi-k2.6: Writing", "The judge is ranking the answers…", the winner) without the answer text being read out word by word, and the focus never leaves the box you are typing in. The cards and the verdict fit panels as narrow as 320 pixels, read right to left, and stay legible in light and dark themes.
• The status words on Compare cards (Connecting, Completed, Failed, Timed out) are now translated; they were English in every other language.
• In the forced light and dark themes, the green, red, gold and accent status colours are now chosen to be readable on the forced background.

Verified release

• Quality, unit, integration, extension-host, browser, dependency, and packaging gates passed for commit 0de8eba09bedae6f292faaa7cecbff17feb3ac3a.
• The attached VSIX was reproduced during this workflow and compared with the committed artifact before publication.

Installation

Download clawai-coding-agent-1.90.0.vsix, run Extensions: Install from VSIX... in VS Code, select the file, and reload the window.