Overview

deepcode is a terminal coding agent that talks directly to DeepSeek / GLM / Kimi: full tool orchestration, permissions, memory, subagents, and workflows — every line of it in your hands, not a wrapper and not a black-box harness.

What makes it different

• Native multi-provider, switch at runtime: one harness covers DeepSeek / GLM / Kimi / self-hosted backends; a dialect adapter layer normalizes each vendor's usage, cache-hit, and thinking tri-state fields, so switching vendors is seamless.
• Cost control for Chinese thinking models: thinking: disabled by default saves roughly 39× on output tokens, and fixes a real bug where, on short answers or gated scenarios, Chinese thinking models let reasoning tokens crowd out content.
• Built-in memory system: per-project memory + a cross-project global drawer + background "dream" consolidation + SearchMemory full-text search (node:sqlite FTS5, zero dependencies), with signal-gated extraction that only fires when there's something worth keeping — unobtrusive, proactive recall.
• Reproducible eval harness: anti-contamination scenarios × multiple models × N seeds × programmatic scoring, one command producing pass^N reliability plus cost-Pareto regression reports, turning "reliability per yuan of a cheap model" into a first-class metric.
• Chinese multimodal support: GLM-4.6v vision pass-through and GLM-OCR document input, injected into the main loop model-agnostically.
• Permissions & security: an allow/ask/deny three-bucket model + an auto-mode classifier + layered settings + SSRF protection + git-worktree-isolated subagents — let it run, you decide the boundaries.

Who it's for

Developers who want to run Chinese LLMs from the terminal, with the experience and full control you'd expect from top-tier closed-source agents.

Owning, Not Renting

DeepSeek and others offer compatible endpoints — two environment variables and you can point a top-tier closed-source agent at a Chinese model. But that's renting: the harness is a black box, its system prompt and tool descriptions were tuned for a different model, and the compatibility layer drops fields along the way. deepcode is owning:

	Compat endpoint + closed-source agent	deepcode
System prompt / tool descriptions	Tuned for a different model	Written for Chinese models, editable line by line
Compatibility layer	Drops fields, translation loss	Direct native OpenAI-compatible endpoint, no translation
thinking cost	Determined by the agent's own behavior	Explicitly disabled by default (saves ~39× output tokens)
Multi-provider	Single provider	DeepSeek / GLM / Kimi / self-hosted, switch at runtime
Modifiability	Not modifiable	Every line is yours

Next steps

• Quickstart: up and running in 30 seconds.
• How it works: the main loop, tools, permissions, and memory architecture.