AIClaw v1.10.5: Configurable Web Search for Tool-Using Agents

AIClaw just shipped two updates that make the agent runtime more reliable and more useful for real-world research workflows.

Project: https://github.com/chowyu12/aiclaw

The short version:

• v1.10.4 hardens Runtime Plan State so agent plans behave more predictably across tool failures and final message saves.
• v1.10.5 adds configurable web search engines, including Tavily, SerpAPI, and Aliyun IQS.

This is a meaningful step for AIClaw because search is no longer treated as a hidden model-side feature only. It can now be configured, tested, observed, and audited as part of the agent runtime.

Why this matters

AI agents often need fresh information: latest releases, pricing, recent policy changes, news, documentation updates, and project status. Previously, this usually depended on either model-native search support or ad hoc browser/web fetch behavior.

AIClaw now supports two web search modes:

1. Built-in model search
2. External search engines exposed through a web_search tool

For models that support native web search, AIClaw can send the model request with search enabled. For external search, AIClaw exposes a real tool call and routes it through the selected search engine configuration.

That difference matters. External search becomes visible in execution logs, streams back like other tools, records input/output/duration/errors, and can be managed from the console.

New Search Engine console

The web console now includes a Search Engine page. You can create multiple search configurations and enable or disable them independently.

Supported providers:

• Tavily
• SerpAPI
• Aliyun IQS

The UI supports:

• creating and editing search engine configs;
• storing provider, base URL, API key, and enabled status;
• testing an existing search engine from the list;
• testing a new or edited config before saving;
• showing whether an API key has been configured without exposing the key value.

For Aliyun IQS, AIClaw calls:

POST https://cloud-iqs.aliyuncs.com/search/unified

with Bearer API key authentication and the LiteAdvanced engine.

Agent-level search mode

Agents now have web search settings in the Agent form.

Each agent can choose:

• web search disabled;
• built-in search, when the selected model supports it;
• external search, using one enabled Search Engine configuration.

If an agent uses external mode, it must select an enabled search engine. Disabled search configs can still be tested from the Search Engine page, but they are not available during live agent execution.

This keeps runtime behavior explicit: the agent does not silently use an unknown search backend.

A real web_search tool

AIClaw now includes a built-in web_search tool for external mode. The backend normalizes search requests, limits result count, trims long snippets, and returns structured search results:

{
  "query": "...",
  "provider": "tavily",
  "results": [
    {
      "title": "...",
      "url": "...",
      "snippet": "..."
    }
  ]
}

The implementation includes provider-specific adapters for Tavily, SerpAPI organic results, and Aliyun IQS page items.

Search observability

Search is also part of the execution trace now.

For built-in model search, AIClaw records a web_search execution step showing the search input and request configuration. For external search, the web_search tool call is tracked like other tools, including output, duration, and errors.

That makes search behavior easier to debug. If an answer depends on web results, you can inspect how the search was executed instead of treating it as a black box.

Plan State hardening in v1.10.4

The previous patch release focused on Runtime Plan State reliability.

Key changes include:

• stronger validation for plan item states;
• duplicate plan item ID rejection;
• normalization so only one item can be running;
• safer failure transitions when a tool or LLM call fails;
• final message linking that completes or fails active plan items correctly;
• compact Plan State prompt text truncation to avoid injecting oversized plan fields into the model context.

This matters because AIClaw treats plans as runtime state, not as chat-visible TODO text. The harness owns lifecycle transitions, persistence, streaming, and failure recovery. The new tests make that contract harder to accidentally break.

Tests added

The update adds coverage for:

• Tavily search parsing;
• SerpAPI search parsing;
• Aliyun IQS search parsing;
• disabled or missing search engine configuration;
• selected search engine routing;
• built-in search request configuration;
• web_search execution step recording;
• plan state normalization, validation, failure handling, and final message linking.

Install or update

AIClaw publishes prebuilt binaries for Linux and macOS:

curl -fsSL https://raw.githubusercontent.com/chowyu12/aiclaw/master/install.sh | bash

Common commands:

aiclaw start
aiclaw stop
aiclaw status
aiclaw update
aiclaw version

What is AIClaw?

AIClaw is a self-hosted AI agent platform for building, operating, and observing tool-using agents. It combines a Go backend, Vue 3 admin console, multi-provider model support, runtime Plan State, tools, skills, sub-agents, memory, and messaging-channel integrations in one deployable binary.

GitHub: https://github.com/chowyu12/aiclaw

If you are working on agent runtime design, tool execution, MCP, Plan State, or observable web search, feedback and issues are very welcome.