Dev Tools
Evaluate and integrate any API with Omnigent and Nimble
Three Omnigent agents turn an API shortlist into an auditable decision and a mergeable integration: a Nimble Web Search Agent researches each candidate against one schema with per-field citations and confidence, striking any unsourced claim before it can sway the outcome. A Claude head and a GPT head then argue over that evidence, a coding agent implements the winner, and a different vendor's model reviews the diff.
Quick Start
Inputs
- A shortlist of candidates The APIs, libraries or vendors you are choosing between. The worked example uses geocoding providers, but the list is a parameter.
- The facts that decide it One schema every candidate is measured against: pricing, limits, licence terms, whatever your decision actually turns on.
- Two model vendors One head on Claude, one on Codex. Existing CLI logins are enough.
Outputs
- A cited comparison across every candidate, one schema
- Per-field citations and confidence
- A detailed decision from two vendors models
- An explicit list of what the evidence could not settle
- A working integration on a reviewed branch
How it works
A three-act chain, each act handing the next one a file.
- Build the evidence once One Nimble Web Search Agent run per candidate against a shared schema, sourced from each vendor's own pages.
- Cite per field, not per paragraph Claims come back keyed to schema fields, each carrying its own sources and confidence.
- Two vendors, one file A Claude head and a GPT head read the same evidence, so they argue about judgment rather than facts.
- Strike what is not sourced An uncited claim cannot decide the outcome, and a conditional permission stays conditional.
- Treat disagreement as a signal Where the two heads diverge is usually where the evidence has a hole.
- Build it, then have another vendor check A coding agent implements the winner in its own worktree; a different vendor's model reviews the diff.
Stack
Nimble primitives plus the full runtime stack.
Nimble APIs
What it does
- Web Search Agents One research run per candidate, returning a schema-conforming answer with per-field citations and confidence.
3rd Party Tools
Role
- omnigent 0.9.0+ The meta-harness from Databricks. Install with the nimble extra.
- Claude Code CLI Runs one decision head and one coding worker, off a subscription.
- Codex CLI Runs the second head, the second worker, and the evidence agent.
- tmux Required by Omnigent's native harnesses.
- python 3.12+ Required by omnigent.
Reach out if you have any questions.
Talk to an Expert