Context Providers

Context providers inject dynamic runtime data as system messages into the LLM context on every call. Unlike tools (which the model calls on demand), providers run automatically before every LLM request.

Use cases

• Server's own public URL (built-in: public_url)
• Current hostname or environment name
• Active feature flags
• Short status summaries that should always be visible

How it works

On every LLM call, Agent._collect_context_injections(profile) runs all active providers and returns their output as role="system" messages. These are inserted right after the memory summary, before conversation history.

Which providers run for a given call:

1. All providers where global_provider = True (always, regardless of profile)
2. Providers listed in the profile's context_providers config field

Built-in providers

User providers

Drop a .py file in context_providers/ at project root. Call reload_tools to activate.

Module format

name = "my_provider"
description = "What this injects."
global_provider = False  # True = all profiles, False = opt-in per profile

async def get_context() -> str | None:
    return "[System] ..."  # or None to skip

Template

See context_providers/_template.py for a copy-paste starting point.

Profile configuration

Add provider names to context_providers in config.json to opt in to non-global providers:

{
  "context_providers": ["hostname", "active_project"]
}

Default is an empty list — only global providers run.

Reload

reload_tools reloads both user tools and user context providers in one call. Built-in providers (in navi/context_providers/) are loaded at startup only.

Writing providers (Navi)

Use tool_manual("write_context_provider") for the full format reference and examples before writing a provider file.

File locations