Build, distill, and run document verification systems with an LLM agent. Pure Node.js. One binary. Bring your own model.
KC is a coding agent purpose-built for rule-based document verification: read a regulation, decompose it into atomic verification rules, write skills to check each rule against sample documents, and (optionally) distill those skills into cheap worker-LLM workflows for production batch processing.
It is designed for the developer at a bank, insurer, or law firm who needs to verify hundreds of documents against dozens of compliance rules — and wants the system to be transparent, testable, and ownable.
npm install -g kc-beta
kc-beta onboard # configure provider + API key
cd my-project # a folder containing rules/ and samples/
kc-beta # launch the agentRequires Node.js 20+. See QUICKSTART.md for the full setup walkthrough.
KC drives a single coding agent through seven phases:
| Phase | What it does |
|---|---|
| BOOTSTRAP | Set up the workspace, detect rules/samples in your project |
| EXTRACTION | Decompose regulation documents into atomic, testable rules |
| SKILL_AUTHORING | Write a verification skill for each rule (Anthropic skill-creator format) |
| SKILL_TESTING | Run skills on samples, iterate via the evolution loop |
| DISTILLATION | Convert proven skills into cheap worker-LLM workflows |
| PRODUCTION_QC | Run workflows on production batches with confidence-based sampling |
| FINALIZATION | Package deliverables — canonical layout, README, coverage report, final dashboard (v0.6.0) |
The conductor LLM (your main model) drives all reasoning. Worker LLM tools are gated to DISTILL + FINALIZATION phases, so the build phase is always grounded in high-quality output.
~/.kc_agent/
config.json # provider, API key, model tiers
workspaces/<sessionId>/ # KC's working files
rules/, rule_skills/, workflows/, samples/, output/, logs/
AGENT.md # per-project context (KC can edit)
tasks.json # ralph-loop task list
session-state.json # phase + pipeline state for /resume
your-project/ # where you launched kc-beta
rules/ # source regulations (KC reads with scope="project")
samples/ # sample documents
Output/ # KC writes user-facing reports here
Dual-directory design. KC has full read/write to its own workspace plus
scoped read/write to your project directory. Source files stay in your
project; KC's working artifacts stay in ~/.kc_agent/workspaces/.
Phase-gated tools. Worker LLM, workflow runner, tier downgrade, and QC sampling tools only register during DISTILL phases. BUILD phases force the conductor to do the intellectual work directly — the results are the ground-truth baseline for distillation.
Skills as first-class deliverables. Every rule produces a self-contained skill folder (SKILL.md + scripts + references + samples). For complex rules that worker LLMs can't reliably handle, the skill itself — run by a capable agent — is the production solution.
10 providers configured out of the box:
- SiliconFlow (default, recommended for China)
- Aliyun Bailian (with coding-plan key support)
- VolcanoCloud (ByteDance Doubao)
- Anthropic (Messages API native)
- OpenAI
- Zhipu GLM
- MiniMax
- OpenRouter
- AWS Bedrock (stub)
- Custom (any OpenAI-compatible endpoint)
Model assignments live in src/model-tiers.json —
edit directly to update tier-1 through tier-4 LLM and tier-1 through tier-3
VLM (vision) models per provider, no code changes needed.
You can use separate providers for the conductor and worker LLMs (e.g., Anthropic conductor + SiliconFlow workers).
When KC extracts rules, it automatically generates a per-rule task list and processes them one at a time. Between tasks the conductor's context is compacted aggressively, so context stays bounded even with 50+ rules.
SKILL_AUTHORING [████████░░░░] 8/12
✓ R001 Registered capital check
✓ R002 Net asset adequacy
▸ R003 Related-party disclosure ← current
· R004 Risk capital calculation
...
Use /tasks to see the full list. The agent decides how to do each task;
the task manager only tells it what's next.
/help Show available commands
/status Session, model, phase, context usage, parallelism
/tasks Show task list and progress
/tools List registered tools + which phase gates each (v0.6.0)
/phase [sub] advance | status | <name> — manual phase override
/parallelism [N] Show / set parallel ralph-loop worker count 1-8 (v0.6.0)
/schedule Show scheduled ingestion jobs
/clear Clear conversation (workspace preserved)
/compact Summarize older messages via the conductor
/sessions List all sessions
/resume <name> Resume a previous session
/rename <name> Rename current session
/exit Save state and quit
--en / --zh / --parallelism=N flags override for one session without writing config.
v0.6.0 TUI upgrades: input stays active during streaming (type-ahead queues automatically), left/right arrow keys move the cursor, up/down recalls session history, Ctrl-A/Ctrl-E jump to start/end. CTX status bar smooths over 30 samples and tracks a session peak.
Some heavyweight features ship as meta-meta skills the agent invokes on demand, rather than always-on dependencies:
pdf-review-dashboard— Two-column HTML dashboard (PDF on the left, verification results on the right, click-to-jump) for manual review and ground-truth collection.auto-model-selection— Use Context7 CLI to fetch current model listings when the bundledmodel-tiers.jsonis stale or you've switched providers.
Both are bundled in template/skills/{en,zh}/meta-meta/ and discovered by
the skill loader at startup.
--parallelism=N (1-8) runs up to N verification tasks concurrently via
subagent orchestration. Safety guardrails:
- Silently clamps to 1 unless
KC_PARALLELISM_VERIFIED=1is set in the workspace.env— prevents accidental multi-hundred-dollar runs before you've confirmed heap behavior. logs/heap.jsonlis sampled every 60 s permanently. Runnode scripts/heap-analyze.jsto get a FLAT / DRIFTING / GROWING verdict from the current workspace's run.rules/catalog.jsonwrites serialize through a POSIX file lock; concurrent workers no longer race each other.agent_tooloperations:spawn/wait/poll/list/killgive the agent visibility and control over its own subagents. Astale_subagentspipeline event on phase_advance lets the main agent clean up before moving on.
Recommended flow: run a 2h serial baseline first, confirm the heap
verdict is FLAT, then set KC_PARALLELISM_VERIFIED=1 and try N=2.
Global config: ~/.kc_agent/config.json (set by kc-beta onboard).
Per-project override: <project>/.env.
Edit anytime with the category-based editor:
kc-beta configCategories: LLM Provider, Model Tiers, VLM Tiers, Worker LLM Provider, Quality Thresholds, Language.
- QUICKSTART.md — full setup and slash command reference
- DEV_LOG.md — release history and design rationale
- docs/global_update_design_v3.md — v3 design plan and progress tracker
- docs/initial_spec_draft.md — original architectural spec
v0.7.3 — codex review patch release. Latest line in the v0.7.x hardening track. Architectural payload from v0.6.0+ is still in place:
- Parallel ralph-loop (up to 8 concurrent workers) with a heap-safety conformance gate
- Native chunker + RAG (onion-peeler + CJK bigram keyword index + one-shot LLM bundle classifier, ported from the AMC verification app)
- Agent-owned task board: the agent reads the rule list from
describeState, decides decomposition (per-rule / grouped / range), and callsTaskCreate/TaskUpdate/TaskCompleteto drive the Ralph loop. Source-context auto-attach pulls rule NL + evidence chunks- sibling rules into the prompt of each task as it runs.
- Workspace file locking for shared coordination files (
rules/catalog.json,rules/manifest.json,refs/manifest.json,tasks.json,session-state.json) — every writer goes throughwithFileLock. agent_toolgetswait/poll/list/killoperations +stale_subagentsphase-advance signal- FINALIZATION phase packages the session into a shippable deliverable
(canonical
rule_skills/layout + README + coverage report + final dashboard) - Filesystem-derived phase milestones (v0.7.0+): the engine reads disk artifacts for advance criteria, never trusts tool-call assertions
- Input stays active during streaming (type-ahead queue), arrow keys +
history recall, CTX smoothing + peak, per-provider context-limit caps,
/tools,/parallelism, and more
See DEV_LOG.md for the per-release change breakdowns and docs/update_design_v7.md for the v0.7.x plan and patch notes.
Bug reports and PRs welcome at https://github.com/kitchen-engineer42/kc-cli.
KC v0.7.0+ is dual-licensed under PolyForm Noncommercial 1.0.0 plus a separate commercial license available on request.
- Personal users, students, hobbyists, public-research orgs, charities, and government institutions — use, modify, and self-host KC for free under LICENSE (PolyForm Noncommercial 1.0.0).
- Enterprises in production (for-profit company workflows, hosting KC as a paid service, distributing KC inside a commercial product) — require a commercial license. See LICENSE-COMMERCIAL.md for terms and how to contact us.
- Redistribution as a competing or independent product — forbidden
under both license tracks. Internal forks for licensee use are fine
with the
Required Notice:preserved; releasing KC under another name as a new offering is not.
KC v0.6.x and earlier remain under MIT for those release versions (licenses can't be retroactively changed); the v0.7.0 cutover applies to all subsequent commits and releases.
Bundled meta-skills under template/skills/ follow the same dual
license as KC itself.
Built by Memium / kitchen-engineer42.