agentic-fabric

Framework-agnostic agent fabric orchestration. Declare fabric agents once in YAML, then run them on CrewAI, LangGraph, Strands, or local CLI runners without changing a single fabric agent definition. Runtime frameworks are optional and are detected lazily from what is installed.

Documentation | Source | Issues

Installation

# Core discovery, loading, runner selection, local CLI, and neutral file tools
pip install agentic-fabric

# With a specific framework
pip install "agentic-fabric[langgraph]"
pip install "agentic-fabric[strands]"

# Non-framework optional surfaces
pip install "agentic-fabric[a2a]"
pip install "agentic-fabric[mcp]"
pip install "agentic-fabric[scraping]"

# Add only the provider SDKs the agent actually calls
pip install "agentic-fabric[github,slack]"

Local CLI runners do not require a Python extra. Install the external CLI (aider, claude, codex, ollama, or a custom executable) and inspect profiles with agentic-fabric list-runners --json.

vendor-fabric is a required dependency and owns provider routing. The anthropic, aws, cursor, github, google, meshy, secrets-sync, slack, vault, and zoom extras pass through to its matching optional dependencies. Provider SDK imports remain lazy.

There is no aggregate AI extra. Install exactly the framework or provider path you use. The CrewAI adapter remains lazy, but agentic-fabric does not publish a CrewAI extra while CrewAI depends on ChromaDB releases covered by an upstream critical advisory with no patched version. Install CrewAI separately only after reviewing that advisory state. Core, local-CLI, and first-party scraping installs are unaffected.

Quick Start

1. Define a Fabric Agent (YAML)

# .fabric/fabric_agents/analyzer/agents.yaml
code_reviewer:
  role: Senior Code Reviewer
  goal: Find bugs and improvements
  backstory: Expert at code analysis

# .fabric/fabric_agents/analyzer/tasks.yaml
review_code:
  description: Review the provided code for issues
  expected_output: List of findings with severity
  agent: code_reviewer

2. Run It

from pathlib import Path

from agentic_fabric import detect_framework, get_fabric_agent_config, run_fabric_agent_auto

# See what framework is available
framework = detect_framework()

# Load a fabric agent manifest discovered in a package or workspace
config = get_fabric_agent_config(Path(".fabric"), "analyzer")

# Auto-detect best framework and run
result = run_fabric_agent_auto(config, inputs={"code": "..."})

Or from the CLI:

agentic-fabric run my-package analyzer --input "Review this code: ..."

3. Use a Specific Runner

from agentic_fabric import get_runner

runner = get_runner("langgraph")  # Force LangGraph
fabric_agent = runner.build_fabric_agent(config)
result = runner.run(fabric_agent, inputs)

4. Carry Runtime Context with Data

from agentic_fabric import AgenticData, get_framework_info

print(get_framework_info())

session = AgenticData({"repo": "jbcom/agentic-fabric"})
session.register_fabric_agent("reviewer", config)
result = session.run_fabric_agent("reviewer", runtime="crewai")

Key Features

• Framework agnostic: one fabric agent definition, multiple runtime backends.
• Lazy imports: core package import does not require CrewAI, LangGraph, Strands, or vendor SDKs.
• Framework-neutral file tools: built-in filesystem tools can be resolved without installing CrewAI or Pydantic; framework adapters add schema wrappers only when their optional dependencies are present.
• Focused extras: langgraph, strands, a2a, mcp, scraping, tests, typing, docs, and dev.
• AgenticData: carries data, registered fabric agents, active runtime selection, and vendor-layer context together.
• Capability decorators: runners and tools expose declared capabilities through read-only metadata and deterministic dispatch.
• Tool resolution: built-in, vendor URI, and registered factories are preferred; external dynamic imports require AGENTIC_FABRIC_TOOL_IMPORT_ALLOWLIST.
• Vendor tool catalogs: AgenticData.vendor_tools() adapts inherited VendorData capability metadata into agent-facing tools without importing provider SDKs directly.
• YAML-first: fabric agent configuration in YAML, not Python boilerplate.
• Hierarchical orchestration: ManagerAgent delegates across fabric agents.
• Package discovery: finds .fabric/, .crewai/, .langgraph/, and .strands/ directories.
• Vendor passthrough extras install the matching vendor-fabric provider dependency without moving connector behavior into this package.
• CLI and library: use from the command line or import as a module.

Framework Priority

1. CrewAI, if installed.
2. LangGraph, if CrewAI is unavailable.
3. Strands, if neither CrewAI nor LangGraph is available.

You can always force a specific runner with get_runner("langgraph") or agentic-fabric run --framework langgraph.

If the selected runtime is not installed, errors point to the matching agentic-fabric[...] extra. Framework-specific config directories also enforce their runtime: a fabric agent in .langgraph/ will not silently run on CrewAI.

Local CLI Runners

For single-agent coding tools, use the --runner CLI path:

agentic-fabric list-runners --json
agentic-fabric run --runner aider --input "Add validation to auth.py"
agentic-fabric run --runner ollama --model deepseek-coder --input "Explain this module"

Profiles are loaded from the packaged local_cli_profiles.yaml, validated before use, and rejected on POSIX systems if the profiles file is group- or world-writable.

A2A Interfaces

The a2a extra installs the official A2A Python SDK and Starlette/Uvicorn HTTP surface. Agentic Fabric implements A2A Protocol 1.0 with Agent Card discovery and the JSON-RPC binding:

from agentic_fabric import create_fabric_agent_spec, create_a2a_app

spec = create_fabric_agent_spec("reviewer", "https://agents.example.com/a2a")
app = create_a2a_app(spec)

create_vendor_a2a_app() exposes a single structured skill accepting provider, operation, and arguments; execution delegates to AgenticData.call() and therefore to VendorData. The bundled development entry point is:

agentic-fabric-vendor-a2a --host 127.0.0.1 --port 8000 \
  --url https://agents.example.com/a2a

The app uses an in-memory task store and declares streaming but not push notifications. Wrap it with application authentication, authorization, rate limits, audit logging, HTTPS, and a durable task store before production use. Only JSON-RPC is advertised; HTTP+JSON/REST and gRPC are not silently claimed.

MCP Adapters

The mcp extra installs the stable MCP Python SDK 2.x, which implements the 2026-07-28 protocol while retaining earlier-client compatibility. It enables two stdio console entry points:

agentic-fabric-vendor-mcp
agentic-fabric-meshy-mcp

agentic-fabric-vendor-mcp exposes credential-free public catalog functions and currently available VendorData.capabilities(). Calls route through AgenticData.call(); the adapter never imports connector implementations or uses Vendor Fabric's private registry. agentic-fabric-meshy-mcp converts Meshy capability metadata from vendor-fabric[meshy] into MCP tools. Both servers validate tool inputs against Draft 2020-12 JSON Schema and return both structured content and a serialized text fallback. Declared output schemas are also enforced. Provider results and diagnostics pass through Extended Data's redaction helpers before leaving the process.

The generic MCPToolAdapter and create_tool_server() API can expose an application-owned callable set without a vendor dependency. The shipped CLIs use stdio; applications can mount the returned SDK server over another MCP 2.x transport when they own the transport security policy.

Repository Boundary

• extended-data owns base data containers, logging, input handling, files, redaction, and generic workflows.
• vendor-fabric owns vendor connectors, provider-backed sync, the SecretSync Python facade/capability surfaces, provider capability metadata, and provider dispatch.
• agentic-fabric owns fabric agent discovery, runner selection, framework adapters, A2A/MCP protocol surfaces, agent-facing tool wrappers, and orchestration.

Protocol behavior follows the current A2A 1.0 specification and MCP 2026-07-28 tool specification.

Full guides and API documentation are published at jonbogaty.com/agentic-fabric. AGENTS.md contains durable repository guidance for Codex sessions.

Documentation

The docs are built with Sourcey. The public API-reference Markdown is deterministically generated from the two package __all__ surfaces:

python scripts/generate_api_reference.py --check
npm ci --prefix docs
npm run build --prefix docs

Local validation:

uv sync --all-packages --all-extras --dev
tox -e lint
tox -e typecheck
tox -e audit
tox -e py311
tox -e py312
tox -e py313
tox -e py314
tox -e coverage
tox -e plugin
tox -e examples
tox -e build

License

MIT