Build software you can trust.
Fast.

Capacitor records and evaluates the process behind the work — turning mistakes into guardrails so agents get smarter, judgement gets easier, and code ships faster with every session.

Get Started
npm install -g @kurrent/kcap
Read Docs

Works with Your Favorite Coding Agent

PiGeminiOpenCodeClaudeCodexCursorKiroCopilot

Your agent’s code gets reviewed.
But its process doesn’t.

A good outcome doesn’t mean a sound process. The scariest result isn’t code that fails, it’s code that looks right but hides things.

01

Safety

Did the agent introduce risk—regressions, vulnerabilities, unintended blast radius?

Coding agents can run commands and modify production code with real consequences. Safety evaluations identify risky behavior before it reaches production, helping teams catch vulnerabilities, destructive actions, and unintended side effects.

Coding agents drop tables they were never asked to touch.
02

Plan adherence

Did the agent do exactly what was asked — no more, no less?

Agents drift. A login fix turns into an auth rewrite; a color change turns into a five-file reorganization. Plan Adherence evals verify the agent stuck to the ask — and flag it when scope, drift, or a quietly abandoned step slips through unnoticed.

Coding agents turn a 10-line fix into a 200-line refactor.
03

Quality

Is the agent’s code not just functional, but “good” — correct, maintainable, properly tested?

Tests passing isn’t the same as code being good. Quality evals check whether the agent actually tested its own work, handled errors instead of hiding them, and followed the conventions already in your codebase — the things a green CI check won’t tell you.

Coding agents write happy-path tests and call it done.
04

Efficiency

Was this the simplest path to the outcome, or was it bloated?

Working code can still hide a wasteful path to get there — retries, redundant calls, entire builds that had to be ripped out and redone. Efficiency evals measure the cost of how the work happened, not just whether it ultimately worked.

Coding agents retry a failing command without changing anything.

Your software factory, instrumented.

What you can see, you can fix. Drive the actions that matter.

INSIGHTS

Developer scores

4.3
INSIGHTS

Strong run. Efficiency is the drag.

17 of 20 points across 13 questions · 3 pass, 1 warn · above the 50-session average of 4.0

SAFETY
5
of 5
pass
avg 4.8 · 4 questions · no findings
QUALITY
5
of 5
pass
avg 4.6 · 3 questions · no findings
PLAN ADHERENCE
4
of 5
pass
avg 4.0 · 3 questions · 1 tip
EFFICIENCY
3
of 5
warn
avg 3.1 · 3 questions · 2 tips
ACTION

What to change next run

Sorted by impact across last 50 sessions
01 Add retry backoff to network-dependent tool calls Efficiency 3/5
02 Escalate environment failures instead of debugging them Efficiency 3/5
03 Resolve label conflicts before creating follow-up issues Plan adherence 4/5
Re-evaluate Session Session is running again. Re-evaluation is available once it ends.
Repeat patterns
9 findings · last 50
Redundant tool calls after failure 12 / 50
Long environment debugging detours 8 / 50
Incomplete follow-up issues 6 / 50
MEMORY.md edited without re-read 4 / 50
rm -rf outside temporary paths 1 / 50

Create your Institutional Knowledge Base

Capacitor creates a software factory that remembers, learns and improves with every session.

01.

Session

the session trace (JSONL)

Agent session ends. Full trace captured — tool calls, errors, corrections, causal chain.

02.

Judge

13 judges · no tools

Per question: pass / warn / fail + evidence. Per category: 1-5 aggregate.

SafetyPlan AdherenceQualityEfficiency
03.

Retrospective

human-readable output

Suggestions · Issues · Strengths

Retrospective cardPer-category cards
04.

Fact Extracted

Facts tab · per repo

Each finding becomes a fact.

mutedeletesuppressed
05.

Clustering

embedding dedup

Facts cluster into patterns.

weightage decayN phrasings
06.

Curation

Curation tab · per repo

Promote or dismiss?

promotedismissregressions
07.

Injected

guardrails

Curated guidance feeds the next session.

Injection · ephemeralWrite-back · durable
01.

Session

the session trace (JSONL)

Agent session ends. Full trace captured — tool calls, errors, corrections, causal chain.

02.

Judge

13 judges · no tools

Per question: pass / warn / fail + evidence. Per category: 1-5 aggregate.

SafetyPlan AdherenceQualityEfficiency
03.

Retrospective

human-readable output

Suggestions · Issues · Strengths

Retrospective cardPer-category cards
04.

Fact Extracted

Facts tab · per repo

Each finding becomes a fact.

mutedeletesuppressed
05.

Clustering

embedding dedup

Facts cluster into patterns.

weightage decayN phrasings
06.

Curation

Curation tab · per repo

Promote or dismiss?

promotedismissregressions
07.

Injected

guardrails

Curated guidance feeds the next session.

Injection · ephemeralWrite-back · durable
loops back to 01 · Collected

Dial in your AI-native software factory

Software Factory Functions

Enables the capabilities teams use to build and trust software faster.

Shared Session Memory Layer

Stores session events natively in unique event streams.

  • Immutable – the exact record, as it happened. Nothing overwritten, nothing lost.
  • Ordered – events are captured in the exact sequence they occurred
  • Indexed – events are organized for fast lookup and filtering
  • Replayable – the full sequence of events can be re-run or re-processed at any time

Session Capture

Records every session event across coding agents.

Start building with Capacitor

Get Started Read Docs