fipsagents 0.22.0

Production-ready AI agent framework for FIPS/OpenShift environments

fipsagents

Production-ready AI agent framework for FIPS/Red Hat AI environments. Provides BaseAgent — a pure Python, async-first base class that handles LLM communication, tool dispatch, MCP connections, prompt loading, skill management, configuration, and lifecycle so your agent subclass stays small.

Install

pip install fipsagents[server]

With optional MemoryHub support:

pip install fipsagents[server,memory]

With Prometheus metrics endpoint:

pip install fipsagents[server,metrics]

With OpenTelemetry trace export:

pip install fipsagents[server,otel]

Or vendor the source directly into your project for full control:

fips-agents vendor

See VENDORED marker in src/fipsagents/ for provenance tracking.

Quick start

from fipsagents.baseagent import BaseAgent, StepResult

class MyAgent(BaseAgent):
    async def step(self) -> StepResult:
        response = await self.call_model()
        response = await self.run_tool_calls(response)
        return StepResult.done(response.content)

if __name__ == "__main__":
    from fipsagents.server import OpenAIChatServer
    server = OpenAIChatServer(agent_class=MyAgent, config_path="agent.yaml")
    server.run()

What's included

• LLM client via the openai async SDK — connects to any OpenAI-compatible endpoint (vLLM, LlamaStack, llm-d)
• Two-plane tool system — @tool decorator with agent_only, llm_only, or both visibility
• MCP client via FastMCP v3 — connect to remote servers (tools, prompts, and resources)
• Prompt loading — Markdown with YAML frontmatter
• Skills — agentskills.io progressive disclosure
• Configuration — YAML with ${VAR:-default} env var substitution
• Pluggable memory — memoryhub, markdown, sqlite, pgvector, llamastack, custom, or null. Budget presets (small/medium/large) auto-tune for model context size. Deferred loading, user-turn injection for small models, and per-turn recall patterns
• Protective patterns — max iterations, exponential backoff, rate limiting
• HTTP server — OpenAI-compatible /v1/chat/completions endpoint with streaming, extended sampling parameters (top_p, top_k, repetition_penalty, reasoning_effort), and multimodal content blocks (text + image_url). Image bytes can be referenced inline as data: URIs, by URL, or by the internal file_id:<id> scheme — uploads via POST /v1/files are resolved server-side from the configured BytesStore before the model call
• File uploads — POST /v1/files accepts multipart uploads, persists each file via the configured FileStore, and exposes them either as extracted-text context (pass file_ids: [...] on a chat completion) or as image bytes (reference file_id:<id> from an image_url content block). Opt-in via the [files] extra
• run_tool_calls() — one-line tool dispatch loop for non-streaming agents
• Agent identity — name, description, version exposed via /v1/agent-info

Key methods

Method	Purpose
call_model()	LLM completion with auto-included tool schemas
run_tool_calls(response)	Execute tool calls and loop until the model stops
call_model_json(schema)	Structured output with Pydantic validation
call_model_validated(fn)	Call, validate, retry with backoff
use_tool(name, **kw)	Agent-code tool call (plane 1)
connect_mcp(target)	Connect to an MCP server
get_mcp_prompt(name)	Render an MCP-provided prompt
read_resource(uri)	Read an MCP-provided resource

Used by

This package is the shared framework for templates scaffolded by the fips-agents CLI:

• agent-loop — single-agent loop (step() in a loop)
• workflow — directed graph of nodes with typed state

License

Apache 2.0