Halon is the open-source AI SRE. It investigates production incidents, proves the root cause, and ships the fix.
In every workspace, with $25 of AI included each month. Or bring your own AI key.
Ask Halon in the incident channel, or press Investigate on the incident. You see every step it takes, the evidence behind every answer, and nothing risky happens without a person.
Investigates
Answers
Shows its work
Fixes
Approvals
Halon works from the resource map, the catalog, your handbook, confirmed memory, every past incident and your connected tools. Connections keep it current, and your team decides what it keeps.
Each run shows the theories Halon tested, supported or ruled out, the steps that decided them and what every tool returned.
Every incident leaves Halon better informed. Your team says which lessons are right, the handbook sets the rules, and Halon follows a run, a build or a deploy to the end for you.
Memory
When an incident ends, Halon shares what it learned and your team says which lessons are right. A lesson is a starting point, and live data always wins.
Handbook
Write down how you release, what counts as production and who directs Halon in an incident. Halon follows it every time. When what it sees says otherwise, it suggests an edit, and nothing changes until your team agrees.
Watches
Ask Halon to tell you when a run, a build or a deploy finishes, or to keep checking until a fix holds. You hear one line when something happens, and it changes nothing while it watches.
Reads never wait, safe changes just happen, and anything risky asks the right person. Your keys stay with Firefight, and every call lands in the activity log.
The cloud, logs, code and databases you connect, with only the tools you switch on.
IntegrationsYour keys are stored encrypted, never logged and never shown back in full.
AI accountsEvery call Halon makes goes through Firefight, with no more access than you gave it.
AI data handlingSafe changes just happen. Risky ones ask the right person first.
Changes without askingA fix is written in a throwaway sandbox that holds no keys, and arrives as a pull request.
InvestigationsEvery read and every change is in the activity log, with who Halon was working for.
ActivityHalon reads logs, metrics and code without stopping to ask.
Adding a label or a comment, assigning someone, scaling up within limits, or restarting a service that runs several copies. Each comes with Undo where there is one.
A rollback, a merge or a scale down asks you first. When an approval rule names someone else, it goes straight to them. Nobody is asked twice.
A change to who can reach something, or to a secret, always waits for a person, whatever else you allow.
Halon never holds a key, and secrets in anything it reads are hidden before the AI sees them.
Text in a log, an issue or a web page is something Halon reads, never an order it follows. After reading outside text, Halon asks before any change.
Add your model provider's key and Halon's AI costs go to that account. Without one, Halon runs on Firefight's AI, which every hosted workspace gets each month.
Your own AI account
Add one or more accounts, in the order Halon should try them. If one fails, the next takes over and you are told.
Firefight's AI
On Firefight's cloud your plan includes $25 of AI each month, the trial too. Halon spends it once none of your own accounts can pay. Buy credits when you need more, and they never expire. On your own servers, Halon uses the AI keys your server is set up with.
Every answer here links to the docs page with the detail, and if yours is not here, we reply to every message.
Halon is the open-source AI SRE. It investigates production incidents, proves the root cause, and ships the fix. It works in Slack and in the dashboard, from your resource map, catalog, handbook, memory, incident history and connected tools, and every fact it states links to its source.
Read about HalonAsk in Slack, in an incident channel or by mentioning Firefight, or chat with Halon in the dashboard. Your own AI agents can ask it too.
Starting an investigationOnly what you allow. In a chat Halon has exactly the permissions of the person who asked. Safe changes, such as adding a label or restarting a service that runs several copies, just happen. Anything risky asks first, and a change to access or keys always needs a person. When an alert wakes Halon at night, it fixes something on its own only under rules your team agreed in advance.
What Halon can doEvery answer ships with a confidence level, the facts behind it and the gaps in coverage. Every claim cites the steps it rests on, and a result from a connected tool links back to its page. Your team rates each answer right, partly right or wrong, and the next investigation of the same alert starts from that rating.
How answers hold upNo key does. Halon works through Firefight and never holds a credential. Secrets inside anything it reads are hidden before the model sees them, and keys and tokens posted in a channel are removed before they are stored.
AI data handlingOn your own AI account if you add one, such as Anthropic, OpenAI, Google Gemini, Amazon Bedrock or Azure AI Foundry. Otherwise on Firefight's AI, with $25 of AI included every month on the hosted plan. If you run Firefight yourself, Halon uses the AI keys your server is set up with.
AI accountsYes, as short lessons your team can confirm, correct or delete. When an incident ends, Halon shares what it learned in the channel. It also follows the handbook your team writes, and suggests an edit when what it sees says otherwise.
Halon's memoryAdd Firefight to Slack, connect your tools, and ask Halon on your next incident.
14 days to try every feature, with $25 of AI included. Or run it on your own servers with your own AI keys.
