Self-Describing Capabilities

Agenticore agents auto-discover their enabled features at runtime and inject a capabilities block into every Claude session. This ensures the LLM knows what it can do — voice, Telegram, A2A, observability — without manual prompt maintenance.

Problem

An agent pod may have voice, Telegram, GitHub, and MCP gateway features enabled via environment variables. But Claude doesn’t read env vars — it only sees its system prompt. Without explicit injection, asking “reply by voice” gets “I don’t have voice capabilities.”

How It Works

Pod env vars + Config
        │
        ▼
capabilities.py._build_registry()
        │  reads get_config() + os.environ
        │  builds list[Capability]
        ▼
render_capabilities_prompt()
        │  filters to enabled only
        │  renders markdown block
        ▼
Injected into Claude sessions:
  - Agent Mode: --append-system-prompt in build_claude_cmd()
  - Telegram:   prepended to system message in _process_and_respond()

Capability Registry

Each capability is a simple dataclass:

@dataclass
class Capability:
    name: str        # identifier
    enabled: bool    # derived from env/config at runtime
    description: str # what the agent can DO with this feature

Current capabilities:

Name	Enabled By	Description
`voice`	`VOICE_SERVICE_URL`	STT/TTS via HTTP voice service
`telegram`	`TELEGRAM_BOT_TOKEN`	Connected as Telegram bot
`agent_mode`	`AGENT_MODE=true`	Persistent agent with model/turns config
`agentibridge`	`AGENTIBRIDGE_URL`	Agent-to-agent communication
`redis`	`REDIS_URL`	Session persistence, job queue, events
`observability`	OTel config	OpenTelemetry tracing
`github`	`GITHUB_TOKEN` or GitHub App	Repository interaction
`langfuse`	`LANGFUSE_PUBLIC_KEY`	LLM observability
`litellm_mcp`	`LITELLM_MCP_GATEWAY_KEY`	External MCP tool servers

Injection Points

Agent Mode (CLI sessions)

In agent_mode/agent.py, build_claude_cmd() appends the capabilities block after the system.md system prompt:

from agenticore.capabilities import render_capabilities_prompt

caps_prompt = render_capabilities_prompt()
if caps_prompt:
    cmd.extend(["--append-system-prompt", caps_prompt])

Telegram Connector (in-process)

In connectors/telegram.py, capabilities are merged with TELEGRAM_SYSTEM_PROMPT before building the message list:

from agenticore.capabilities import render_capabilities_prompt

caps = render_capabilities_prompt()
full_system = "\n\n".join(filter(None, [system_prompt, caps]))

Adding a New Capability

One entry in _build_registry() inside agenticore/capabilities.py:

caps.append(Capability(
    name="my_feature",
    enabled=bool(os.environ.get("MY_FEATURE_URL")),
    description="Feature X is available — can do A, B, C.",
))

No other files need changes. Every agent session on pods with that env var will automatically know about the feature.

Output Example

## Agenticore Agent Capabilities

You are **anton-agent**, an agenticore-powered AI agent.

The following capabilities are active on this instance:

- **voice**: You have voice capabilities (speech-to-text and text-to-speech)...
- **telegram**: You are connected to Telegram as a bot...
- **agent_mode**: Running in AGENT_MODE as a persistent AI agent...
- **agentibridge**: Connected to AgentiBridge for agent-to-agent communication...

Use these capabilities when relevant to user requests.
Do not claim you lack a capability listed above.

Design Principles

Zero-maintenance — capabilities derived from existing config. No manual sync when features change.
Declarative — each capability is data, not logic.
Graceful absence — if no capabilities are enabled, render_capabilities_prompt() returns empty string. No injection, no noise.