Reference
What's planned
Babbit is in alpha. These features are designed, and many are partly built, but the deployed product doesn't serve them yet. Where a screen would show one, it shows one line marking it planned.
Why this page exists
Babbit never shows something as done without evidence, and these docs follow the same rule. Rather than describing designed features as if you could use them, the docs mark each one Planned where it comes up, and list them all here.
Build
| Feature | What it will do |
|---|---|
| Branches | A copy of your features, goals and scenarios to change without touching the real product, with its own code branch in each repository, merged as one reviewed change. |
| A fourth goal, placed for you | Choosing whether a fourth goal becomes a child, a merge into the goal it restates or a policy, filing it when sure and asking when not. |
| Placement with confidence | Placing a new feature in the tree with Jev's probability, offering the runner-up, and asking Where does this go? only when unsure. |
| Previews from a workspace | Opening any feature, running, in a drawer at the state a scenario ends in. |
| Rewrite | Rewriting existing code into Babbit's shape in a workspace, scored by how many of your original tests pass. |
| Deepest map levels | Listing every function a step ran, and running an example through them in the browser. |
Manage
| Feature | What it will do |
|---|---|
| Evals | Turning scenarios into graded cases, with a score, a confidence interval and the cost of a run, and grader checks before a number is trusted. |
| Agent evals | Each agent's score over time, with a line per model, from its own past sessions. |
| Climbs | Hill-climbing an agent's prompt, rules or model against an eval, keeping only gains outside noise, behind a token cap. Off unless a project turns it on. |
| Health | The last seven days as a screen: correction loops, done claims without evidence, runs stopped by the runtime, and the slowest and most broken boxes. |
| Skill edits and restore | Editing a skill's text for one project, and restoring Babbit's default or an older version. |
| Speed agent | An agent that starts when a scenario's measure crosses its target, and opens a change to bring it back. |
Platform
| Feature | What it will do |
|---|---|
| Infra screen | Each service on each tier with its version, a person's rollback, restart and run-a-job, and declared replicas. |
| Service page | One page per service: its functions, endpoints, tables, the names of the secrets it reads, and its boxes' numbers. |
| Secrets screen | Listing every secret with its tiers, readers and versions, and adding or rotating one directly. |
| Outside MCP servers | Connecting any MCP server so agents may call its read tools through the Babbit MCP. |
| Usage | Tokens in and out by day and by agent, with workspace time beside them. |
| Migration safety | Refusing expand and contract in one change, and recording a backup on each release before production. |
| Team hand-off | Handing a removed person's open changes and connections to an owner automatically. |