AI Appendix by
Summary and TL;DR
Agentenna is an open-source awareness layer for AI agents. It collects live signals such as messages, logs, alerts, tickets, events, states, and tasks in a local Station, then renders a token-bounded view directly into an agent's context and refreshes that view every turn.
TL;DR: Agentenna gives AI agents a shared, current view of what is happening without putting an unbounded event feed into their context. Producers emit to named channels; each agent reads independently; urgent signals can wake an agent through optional triggers.
Last reviewed
Key facts
- Open source and self-hostable. Agentenna is Apache-2.0 licensed and can run locally with one process and one SQLite database.
- Token-bounded by design. A named awareness slot is replaced each turn, so agents receive a current bounded view instead of an ever-growing transcript.
- Shared channels, independent readers. Many producers can emit to one channel and many agents can listen while maintaining their own read position.
- Signals have distinct lifecycles. Events report what happened, states represent what is true now, and tasks track work that remains open.
- Seeing and waking are separate. Most updates appear on an agent's next turn; optional triggers wake an agent only for signals that cannot wait.
- Built for agent hosts. Agentenna works with Claude Code hooks, MCP hosts, and custom agent loops.
How Agentenna works
A minimal Agentenna flow has three steps.
- 1.
Start a Station. Run the local Station to store signals and serve the Agentenna API.
- 2.
Emit a signal. Send an event, state, or task from a terminal, service, webhook, logger, ticket system, or health check.
- 3.
Connect each reader. Install or integrate Agentenna with an agent host so a bounded awareness block is refreshed in context every turn.
Common use cases
Production incident awareness
Stream errors, deploy state, health checks, and active incident tasks into a coding or operations agent so it can answer with current operational context.
Long-running coding agents
Keep an agent informed about CI results, review feedback, terminal output, and environment changes that occur after its session starts.
Support and ticket escalation
Publish urgent tickets and customer-impacting updates to a shared channel read independently by support, engineering, and coordination agents.
Multi-agent coordination
Let several agents observe the same named channel without point-to-point webhook glue or one reader consuming updates for the others.
Selective agent wake-ups
Keep routine signals passive while attaching triggers to critical faults, urgent tasks, or state changes that require immediate action.
Frequently asked questions
What does Agentenna do?
Agentenna gives AI agents a live, token-bounded awareness feed. It gathers signals from external systems and renders the relevant current view into an agent's context on every turn.
Is Agentenna an agent memory system?
No. Memory helps an agent recall the past. Agentenna is an awareness layer for the present: new events, current states, open tasks, and signals that changed while the agent was not looking.
How does Agentenna avoid filling the context window?
Agentenna writes to a named bounded slot that is replaced rather than appended. The agent sees a refreshed current view instead of an unbounded chronological feed.
Can multiple agents read the same channel?
Yes. Channels are shared buses with independent readers. Each agent maintains its own position, so one agent moving forward does not hide a signal from another.
Does every signal wake an agent?
No. Routine signals become visible on the next turn. Triggers are optional and should be attached only to urgent signals that need to wake a receiver.
Where does Agentenna store data?
The self-hosted Station can store signals in a local SQLite database. The quickstart requires no hosted account, and the landing page states that feeds remain on the user's machine.
Which agent environments can use Agentenna?
Agentenna supports Claude Code hooks, MCP-compatible hosts, and custom agent loops. Producers can emit from terminals, services, webhooks, loggers, ticket systems, and health checks.
Is Agentenna free and open source?
Yes. The Agentenna source is available on GitHub under the Apache License 2.0, and the Python package is published on PyPI.
— The awareness layer for AI agents
Your agent already knows.
Agentenna gives agents a live, token-bounded view of what's happening — messages, logs, alerts, tickets — rendered straight into context, refreshed every turn. Inspect the details, rewind the history, wake up when it matters.
- prod.errors12
- ci.builds5
- terminal.logs28
- support.urgent3
- Billing worker failed processing webhookprod.errors · -2m
- deploy.prod rolled_backstates · -2m
- task opened: investigate checkout 5xxtasks · -11m
One bounded surface
Events, states, and tasks render into a single named slot, replaced every turn. Your agent sees what changed — never an unbounded feed.
Emit once, every agent knows
A channel is a shared bus. Terminals, services, and webhooks emit; every reader tunes in and keeps its own position.
Seeing is not waking
Most signals simply appear on the next turn. Attach a trigger to the few that can't wait, and a receiver wakes up.
Channels
Emit once. Every agent knows.
A channel is a named shared bus, not point-to-point delivery. Many producers emit to it, many readers tune in — and each reader keeps its own position. One reader moving forward never hides the signal from another.
One emit can update all your running agents — without N copies of the same webhook glue.
Quickstart
Emit → Station → Reader.
Python 3.12+, one process, one SQLite file. Everything in the repo is the whole product — no account, no telemetry, and your feeds never leave your box.
pip install "agentenna[station]"
# start a Station — stores signals, serves the API
agentenna serve --db .agentenna/agentenna.db --port 1234
# emit a fact from anywhere
agentenna emit prod.errors "Checkout 5xx rate jumped to 12%" \
--kind fault --severity crit --db .agentenna/agentenna.db
# connect a reader
agentenna install claude-code
Works with Claude Code hooks, any MCP host, or your own agent loop. API cheatsheet · Host guides · Examples