Proxies, Skills, And Sandbox

The v1 proxy, skills, sandbox, and computer-use surroundings are preserved inside native. They are useful for native agents but remain optional and do not reshape the public tactic protocol.

Proxies

A proxy class describes endpoint metadata with BaseProxy.endpoint, and ProxyManager loads registered proxy resources.

from lllm.runtimes.native.proxies import BaseProxy, ProxyRegistrator


@ProxyRegistrator(
    path="demo",
    name="Demo API",
    description="Small local proxy used in examples.",
)
class DemoProxy(BaseProxy):
    @BaseProxy.endpoint(
        category="demo",
        endpoint="lookup",
        name="lookup",
        description="Look up a local value.",
        params={"key": (str, "alpha")},
        response=[{"value": "str"}],
    )
    def lookup(self, params=None):
        params = params or {}
        return {"value": params.get("key", "")}

When an agent has proxy config, native can inject query_api_doc(proxy_name), run_python(code), and a rendered API-directory block into the system prompt.

AgentInterpreter

AgentInterpreter uses exec() in a persistent namespace, injects CALL_API, captures stdout, truncates long output, and returns tracebacks as tool output. Use it for trusted agent workflows and swap the backend when stronger isolation is required.

proxy:
  activate_proxies: [demo]
  deploy_mode: false
  cutoff_date: "2024-01-01"
  exec_env: interpreter
  max_output_chars: 5000
  truncation_indicator: "... (truncated)"
  timeout: 60.0

exec_env: jupyter keeps the API docs but leaves code execution to a tactic using the Jupyter sandbox. exec_env: null injects API awareness without a runtime execution tool.

Skills

Native agent config preserves progressive skill disclosure:

skills:
  - pdf
  - commit-review

Local skills are discovered from standard project and user directories such as .agents/skills/<name>/SKILL.md and .claude/skills/<name>/SKILL.md. A compact catalog is appended to the system prompt, and an activate_skill(name) tool loads full instructions on demand.

skills: "*" exposes all discovered local skills. URL entries are fetched as remote SKILL.md files. Provider-hosted skill ids that start with skill_ are passed through as model args for compatible Anthropic models.

Sandbox And Computer Use

Area	Module	Notes
Jupyter sessions	lllm.runtimes.native.sandbox.jupyter	Notebook file management, kernel execution, markdown/code cells, and optional proxy-aware init code.
In-process interpreter	lllm.runtimes.native.proxies.interpreter	Persistent Python namespace for trusted proxy workflows.
Computer use	lllm.runtimes.native.tools.cua	Browser automation helpers using OpenAI/Azure clients and Playwright when installed.
Response API tools	LiteLLMInvoker	Prompt.addon_args can request web-search or computer-use tools for response API calls.

Install only the extras you need. Offline native tests do not require live providers, Jupyter, Playwright, or proxy API credentials.