## Description Claw Assistant now reads its chat model from `AGENT_MODEL` and falls back to `mastra/openai/gpt-5.4`, so nothing changes until Mastra Platform sets it. This lets Platform's template form offer a model choice (PLTFRM-3312). Web search now uses the Gateway web search tool from docs-expert instead of OpenAI's provider tool, so it works on any model and the agent's instructions stay true. ## Related issue(s) Linear PLTFRM-3315 ## Type of change - [x] New feature (non-breaking change that adds functionality) ## Checklist - [x] I have linked the related issue(s) in the description above - [x] I have made corresponding changes to the documentation (if applicable) - [ ] I have added tests that prove my fix is effective or that my feature works - [ ] I have addressed all Coderabbit comments on this PR Verified: `mastra build` passes and `tsc` is clean against the published packages. <!-- This is an auto-generated comment: release notes by coderabbit.ai --> <!-- coderabbit-ai-generated:v1 --> ## ELI5 Claw Assistant can now use the model named in `AGENT_MODEL`; if the variable is empty or missing, it uses `mastra/openai/gpt-5.4`. Web search uses a separate Gateway tool, so it does not depend on the selected chat model. ## Changes - Documented `AGENT_MODEL` in `.env.example` and the README, including example model IDs and the default. - Added a Gateway web-search tool that sends searches through `openai/gpt-5-mini` with `openrouter:web_search` and returns an answer with deduplicated citations. - Replaced the OpenAI SDK web-search tool with the Gateway tool and removed the `@ai-sdk/openai` dependency. ## Checks - Build and test results: not available. <!-- end of auto-generated comment: release notes by coderabbit.ai --> |
History
|
|
|---|---|---|
| .. | ||
| src/mastra | ||
| workspace/skills | ||
| .env.example | ||
| .gitignore | ||
| package.json | ||
| pnpm-workspace.yaml | ||
| README.md | ||
| tsconfig.json | ||
Claw — Autonomous Personal Assistant
An autonomous assistant that operates a computer to finish multi-step tasks — read/write files, run shell commands in a sandboxed workspace, browse the web with Playwright (@mastra/agent-browser), search the web through Mastra's Gateway, and follow reusable skills checked into workspace/skills/.
Claw has resource-scoped observational memory, so it remembers your preferences and context across threads.
Prerequisites
- Node.js
>= 22.13.0 - A Gateway API key — get one at gateway.mastra.ai
- (Optional) A Turso database; for local dev, the default
file:./mastra.dbworks. - Playwright's Chromium browser for local runs, or
BROWSER_CDP_URLfor a hosted Chrome/Browserbase/Browserless instance in server deployments.
Setup
npm install
cp .env.example .env
# fill in MASTRA_GATEWAY_API_KEY (TURSO_DATABASE_URL defaults to file:./mastra.db)
npm run dev
Then open Mastra Studio (URL printed by npm run dev) and chat with the claw agent.
Workspace layout
workspace/
├── skills/
│ ├── general-tasks/SKILL.md
│ └── research-tasks/SKILL.md
└── … your files end up here
The agent can only read/write inside workspace/ (or whatever CLAW_WORKSPACE_DIR points at). The sandboxed shell runs with that directory as its CWD.
Adding skills
Create a new folder under workspace/skills/<skill-name>/ with a SKILL.md:
---
name: my-skill
description: Short summary the agent uses to decide when to apply this skill.
version: 1
metadata:
tags: [optional, tags]
---
# Process
1. Step one.
2. Step two.
The agent will pick up new skills on the next request.
Environment variables
| Variable | Purpose |
|---|---|
MASTRA_GATEWAY_API_KEY |
Routes the chat model and web search through the Gateway. |
AGENT_MODEL |
Optional model router id. Defaults to mastra/openai/gpt-5.4. |
TURSO_DATABASE_URL |
libSQL/Turso URL. Defaults to file:./mastra.db. |
TURSO_AUTH_TOKEN |
Turso auth token (omit for local file DB). |
CLAW_WORKSPACE_DIR |
Override the workspace root directory. |
BROWSER_HEADLESS |
Set to false to launch the browser headfully. Defaults to headless. |
BROWSER_CDP_URL |
Optional hosted browser CDP URL for server deployments. |
Agent Editor
This template enables the code-backed Agent Editor with new MastraEditor({ source: 'code', codePath: 'mastra/editor' }). Edits made in Studio are written as deterministic JSON overrides under mastra/editor/agents/, so Mastra Platform can open GitHub pull requests for agent changes instead of only saving them to the database.