Completion confidence for coding agents

Install

completion gate

Stop agents from saying "done" too early.

Agents often stop one step early. Homer catches that moment and sends unfinished work back before it lands in your review queue.

Caught before review Blocked
Tests missing

The agent goes back to run them before you read a false "done."

Fewer false finishes. More tasks actually done.

Plugs into:

  • 3 minute install
  • Free usage
  • No credit card
Get install command

Supported by

  • Microsoft
  • Oxford University Innovation
  • IEEE
  • NetMind

demo

The difference is one missing check

Same task, same agent. Homer only changes what happens before "done".

without Homer

Premature finish

missed
Add retry logic and verify the tests pass.
Done. I added retry logic.
evidence tests were not run
User reopens the task after the failing case is found.

with Homer

Completion gated

fixed
Add retry logic and verify the tests pass.
Done. I added retry logic.
Homer check Tests missing

The agent goes back before the answer reaches you.

Runs tests, catches the edge case, fixes retry handling.
Homer check Ready to ship

Evidence supports the final answer.

Get install command

how it works

A completion checkpoint before "done"

  1. 01

    Agent works

    Your coding agent edits files, runs commands, and gathers evidence in the normal loop.

  2. 02

    Homer scores completion

    Before the agent claims the task is finished, Homer returns a calibrated confidence level and explanation from the current evidence.

  3. 03

    Finish or keep going

    Proof ships the answer. Missing proof sends the agent back to fix gaps, verify, and check again.

choose your agent

Install where you already work

Start with one coding agent. Homer uses the same completion confidence API across all three.

Cursor Get Cursor command Claude Code Get Claude command Codex Get Codex command

benchmark

Agent vs Agent + Homer

In-house Terminal Bench 2.0 matched trials analysis - same tasks, same agent, Homer on completion checks only.

Task success

Agent
Agent + Homer

+14% higher with Homer

Premature completion

Agent
Agent + Homer

-61% fewer with Homer

Relative change on matched trials. Premature completion = agent said done while tests or evidence still failed.

Start free install

faq

Common questions

What does Homer actually do?
It scores whether the current user request looks truly complete from the evidence in the turn - before the agent sends a final answer.
Do I need a credit card?
No. Beta install includes free usage. Setup takes about three minutes.
Which agents are supported?
Claude Code, Cursor, and Codex today. Same Homer API and install flow across all three.
Does Homer replace my agent?
No. Homer is a metacognition layer on top - the agent still writes code; Homer gates premature "done".
What happens when confidence is low?
Homer blocks the final answer and gives the agent a concrete reason to keep working, such as missing tests, missing files, or weak evidence.
Does Homer edit code?
No. Homer checks completion evidence. Your coding agent still decides what to edit and what commands to run.
What data leaves my machine?
Homer sends the task, recent transcript, and completion evidence needed for the check. It does not need your editor password or pasted provider API keys.
Can I turn it off?
Yes. Homer installs local settings for supported agents, including friendly controls for enabling checks, strictness, and retries.
Get install command