Typhon
Meet Typhon, your AI teammate.
Typhon is the AI agent at the center of every Gyroscope ticket. A project manager assigns it work; it produces a plan, gets approval through the dashboard, opens a pull request, and exits β the same shape as work assigned to any other teammate.
An agent you assign to, not a tool you reach for.
Most AI coding tools sit beside the developer β invoked, prompted, watched. Typhon sits inside the team. It picks up tickets from Gyroscope's queue the way an engineer picks up tickets from a sprint board. It reads acceptance criteria. It surfaces ambiguities back to the PM when the brief is unclear. It produces a plan, waits for approval, and only then writes code. Project managers can hand it work without ever asking a developer to relay the assignment.
Plan first. Approve. Then implement.
Every Typhon ticket runs in two phases. A read-only session reads the brief, explores the relevant slice of your codebase, and produces a plan. The plan lands on Gyroscope's dashboard for approval. Nothing is edited yet.
Only after approval does a second session pick up the plan as its starting prompt and make the change. The diff that lands in the pull request is exactly what was approved β no archaeology, no surprises.
Built on Claude Code, not around it.
Typhon is a headless Claude Code subprocess at heart. That choice carries weight.
- Every Claude Code feature comes along β skills, slash commands, hooks, plan mode, MCP servers, permission modes, resume.
- Anthropic's Claude Code team owns the agent loop. Tool-use improvements, prompt-cache wins, security work β Typhon inherits all of them, for free.
- Gyroscope handles the orchestration above the agent: which tickets to pick up, what context to feed in, when to ask a human. Typhon handles the implementation. Each layer does what it's best at.
Bring your own Claude.
Typhon doesn't bill you for the tokens it burns. You connect your own Claude credentials, and the agent runs on them.
- Solo developers
- Authenticate with your Claude Max subscription via OAuth. Typhon inherits the credential and burns the same Max quota you already pay for.
- Teams
- Configure an Anthropic API key. Typhon's sessions bill directly to your Anthropic account, with usage visible on Gyroscope's dashboard.
Either way, the credentials never leave your machine, and you only pay Anthropic for the work Anthropic actually did.
Writes its own tools, learns from what doesn't work.
A new Typhon ticket on day 90 should be cheaper and more accurate than the same ticket on day 1 β not because the model got better, but because the agent's accumulated context did.
Tools
Common operations β running a particular query, generating a specific kind of file, checking a recurring invariant β turn into purpose-built tools the agent calls directly. Each one replaces minutes of token-burning improvisation with a single deterministic operation.
Lessons
When an approach fails β a plan that didn't survive contact with the codebase, a fix that broke something else β Typhon notices and writes the learning into permanent memory. The same mistake doesn't get made twice.
Documentation
Every feature ships with the architecture notes that explain it. The next ticket touching that area starts with a fuller picture.
You keep the keys.
Typhon is an agent, not an autopilot. Developers see a live transcript of its work and can veto, redirect, or take over at any point. Project managers approve every plan before code is touched. Reviewers see a PR with a known plan attached β no archaeology required. The agent works fast because the controls around it are tight.