Conifer products

Stable lives inside the coding agent you already use, as a handful of slash commands. You keep your editor, your terminal and your habits. Conifer brings every model, and every other harness, to where you already work.

• /model chooses from any of our 200+ models, or hands the choice to the router.
• /harness runs any harness inside your harness, and you can have several of them running at the same time.
• /reviewer sets any model as the reviewer for any session. It reads what the session writes and answers in the same place.

When a coding session ends, Parallax asks the router for the next best models from other labs and has them go over the work: bugs, security holes, dead code, slow code. Every finding is proved before you see it, so you can ship it with confidence.

• The router chooses the reviewers the same way it chooses anything: the strongest models for the job that did not write the code. Different labs make different mistakes, so they catch each other's.
• A finding counts only when a test fails on your code and passes with the fix, with the rest of your suite still green, in a throwaway copy of your repo.
• The review runs on short, targeted calls, so it costs less than a tenth of the session it reviews. Findings arrive as a pull request, with line annotations on GitHub.

The console is where a team's AI use becomes legible: which models, on which keys, for whom, at what cost, and why each call went where it did.

• Spend by member and by model over the last thirty days, so you can see who is using what, and what it is costing.
• A key per person, per service or per project, minted before you add a card. Revoke one and the gateway stops honouring it within seconds.
• Usage, cost, routing decisions, balance and limits in one place.

Conifer's router decides which model should take each turn. If you build a coding agent, or you run a lot of AI, you can lease that decision and serve the request on your own keys.

• auto and balanced choose the best value among the models that will answer well. best chooses the most capable model, whatever it costs.
• POST /v1/route returns the pick and up to three fallbacks, then gets out of the way. The request itself can go straight to the lab, on your keys, through your stack.
• A named model is always served as that model. The router only decides when you ask it to, and nothing is substituted behind an id you chose.