claudeseek — Agent de IA para Coding Local-First
appleweiping/claudeseek
Agent de coding local com arquitetura industrial, zero deps
Ver no GitHubAgent de coding industrial que roda localmente, combinando a arquitetura do Claude Code com o motor DeepSeek V4. Oferece REPL em terminal com streaming e Web UI local, sem dependências externas. Ideal para automação de tarefas de código com controle total na sua máquina.
Instalação
npx convez-ai appleweiping/claudeseek1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Automação de leitura, busca e edição de código com aprovação manual
- Análise e revisão de segurança em arquivos locais
- Planejamento e execução de refatorações de código
- Execução de agentes de IA sem enviar código para servidores
Veredito

Solução robusta e production-ready para quem quer um coding agent robusto rodando localmente com zero dependências. A arquitetura baseada no Claude Code e suporte multi-provider (DeepSeek, OpenAI-compatíveis) o torna versátil e seguro para uso em ambiente corporativo.
README
A tradução para o português deste README ainda está na fila.
claudeseek is a real coding agent you run on your own machine — not a toy wrapper. It reads, searches, edits, and runs code through a permission-gated tool loop, the same shape Claude Code uses, with two faces over one engine:
- a terminal REPL with streaming output, slash commands, and inline approvals;
- a local Web UI (127.0.0.1, token-gated) with live tool-call cards, an approval flow, a context meter, sessions, and a cost panel.
It reads your machine's DeepSeek credentials automatically — no key to paste — and can switch to any OpenAI-compatible endpoint (a relay that serves Claude / GPT) without touching code. Zero runtime dependencies: just Node ≥ 18.
Lineage. The architecture is studied from the publicly-exposed Claude Code source (
QueryEngine, the tool/permission model, context assembly, the message-centric sub-agent pattern) and re-implemented from scratch in plain ESM JavaScript for the DeepSeek API. It is a sibling to WEIPING_WHALE (the TypeScript DeepSeek CLI) — claudeseek adds the Web UI, the composable permission pipeline, and multi-provider routing. Not affiliated with Anthropic or DeepSeek.
Install
git clone https://github.com/appleweiping/claudeseek.git
cd claudeseek
npm install # installs nothing — there are no dependencies; just links the bin
npm link # optional: puts `claudeseek` / `cseek` on your PATH
node bin/claudeseek.js --doctor
State lives in ~/.claudeseek/ (sessions, undo snapshots, skills, config,
memory outbox). The CLI exposes two commands: claudeseek and cseek.
First run
claudeseek finds your DeepSeek key automatically, in this order:
DEEPSEEK_API_KEY(environment)~/.claudeseek/config.json→llm.api_key- an existing
~/.deepseek-cli/config.toml/~/.weiping-whale/config.tomlinstall (endpoint + key reuse — nothing to re-enter)
claudeseek # interactive REPL
claudeseek -p "summarize this repo's architecture"
claudeseek -m pro --mode plan -p "review auth.js for security gaps"
claudeseek --serve --open # local Web UI
claudeseek --doctor --json # diagnostics, machine-readable
claudeseek --last # resume the most recent session
CLI
The REPL streams the model's thinking and answer, renders each tool call as it runs, and asks before anything risky.
| Slash command | What it does |
|---|---|
/model [name] | show or switch model: pro, flash, auto, reasoner, chat |
/mode [m] | permission mode: default, acceptEdits, bypassPermissions, plan |
/cost | session cost, per-model breakdown, cache-hit ratio |
/context | context-window usage estimate |
/compact | summarize older turns to reclaim context |
/undo | revert the last turn's file changes |
/sessions, /resume <id> | list and reopen past sessions |
/skills, /skill <name> | discover and inject SKILL.md skills |
/todos | show the agent's live task list |
/mcp, /memory | MCP server status · agentmemory status |
/remember <text>, /recall <q> | write to / search agentmemory |
/help, /quit | help · exit (Ctrl+C aborts a running turn) |
--model auto runs a zero-cost keyword router per turn: hard signals
(debug, architecture, 调试, 根因) → V4 Pro + max thinking; light ones
(search, format, 翻译) → Flash; everything else → Flash + thinking.
Web UI
claudeseek --serve --port 3618 --open
A single self-contained page bound to 127.0.0.1, gated by a bearer token
printed once at startup (pin it with CLAUDESEEK_TOKEN). It streams turns over
NDJSON and gives you:
- tool-call cards that expand to show inputs and full output;
- an approval flow — risky calls pop an Allow / Always / Deny card inline;
- a context meter, cost chip, todo panel, and session sidebar;
- markdown + code rendering with copy buttons, light / dark themes;
- one-click model / permission-mode switches, undo, and compact.
The browser talks only to your local server; the server holds the API key.
Providers
claudeseek defaults to DeepSeek (国产 V4 models). To point it at any OpenAI-compatible endpoint — for example a relay that serves Claude or GPT — set one environment variable:
# DeepSeek (default): uses DEEPSEEK_API_KEY
claudeseek -p "..."
# OpenAI-compatible relay: uses OPENAI_BASE_URL + OPENAI_API_KEY
CLAUDESEEK_PROVIDER=relay claudeseek -m gpt-5.5 -p "..."
Under provider: "deepseek" the client sends V4 thinking params
(thinking, reasoning_effort); under provider: "openai" it speaks plain
chat-completions so any compatible backend works. Define named providers in
~/.claudeseek/config.json → providers.
Safety
Permissions follow Claude Code's layered model — first non-null wins:
- hard rules — broadly destructive shell commands (
rm -rf /,Remove-Item -Recurse C:\, fork bombs, disk format) are blocked in every mode; - mode rules —
planis read-only;bypassPermissionsauto-allows; - session grants — "Always allow" remembers a rule for the session;
- kind defaults — reads/search/meta auto-allow; writes ask (auto in
acceptEditsinside the workspace); safe read-only shell prefixes auto-allow; - the approver — REPL prompt or Web UI card. Headless with no approver
denies anything that would need one (use
--yoloto opt out).
This is approval-gating, not an OS sandbox. Writes are snapshotted before
mutation so /undo can roll a turn back.
Architecture
src/
api.js DeepSeek / OpenAI-compatible streaming client (tool-call assembly, retries)
engine.js the agent loop — assemble → stream → dispatch tools → loop → persist
permissions.js composable allow/deny pipeline (hard → mode → session → kind → ask)
context.js system-prompt assembly + project memory (CLAUDESEEK.md / CLAUDE.md / AGENTS.md)
compact.js CJK-aware token estimate + model/offline context compaction
router.js zero-cost per-turn model router (auto mode)
tools/ read_file write_file edit_file list_dir glob grep bash fetch_url todo task
mcp.js stdio MCP client → tools surface as mcp__<server>__<tool>
sessions.js durable JSON sessions (resume by id/prefix, --last)
undo.js per-turn file snapshots
cost.js live cost + DeepSeek prefix-cache accounting
memory.js agentmemory bridge (single source of truth) + offline outbox
repl.js terminal UI
server.js 127.0.0.1 Web UI server (token auth, NDJSON streaming, approvals)
public/index.html the Web UI (zero-build, single file)
Patterns ported from Claude Code: a single streaming agent loop that every
surface consumes; the layered permission decision pipeline; message-centric
sub-agents (results return as <task-notification> the parent reads as
context, not RPC); file-state snapshots for undo; and stable per-session system
prompts for prefix-cache efficiency.
Tests
46 end-to-end checks run fully offline against a mock DeepSeek server — real HTTP, real streaming, real tool execution:
npm test # core + cli + server
npm run test:core # engine, tools, permissions, undo, compaction
npm run test:server # web server: auth, streaming, approval round-trip
Part of the agent workstation
claudeseek is a first-class member of the local agent family documented in
D:\devtools\AGENTS.md. It shares the four-piece kit: Skills (SKILL.md
discovery), Memory (agentmemory at localhost:3111, tagged
agent:claudeseek, per AGENT-MEMORY-PROTOCOL.md), and Safety (the
permission pipeline above). Launch it with claudeseek.cmd.
License
MIT © appleweiping