HagiCode is a proposal-driven AI coding workbench that turns a request into a reviewable proposal, runs it through agent CLIs such as Claude Code, Codex and GitHub Copilot, and keeps the outcome as team knowledge.
DeepAgents is LangChain's open-source agent harness, and its terminal coding agent is now called Deep Agents Code (`dcode`). HagiCode is a workbench that runs it as an executor over ACP. Stay with the CLI for sandboxes, SDK work and scripts; add HagiCode for parallel sessions, proposals and managed commits.
DeepAgents is LangChain's open-source agent harness, built on LangGraph. Its terminal coding agent, which HagiCode calls DeepAgents, is published as Deep Agents Code: install it, run dcode, and it behaves like a coding agent in your terminal. HagiCode is a desktop and web AI programming workbench that drives agent CLIs, DeepAgents among them, as its executor. It does not replace the CLI; it launches it in ACP mode and adds sessions, proposals and commit handling.
If you build your own agents on the Deep Agents SDK, want remote sandboxes or LangSmith tracing, or only need a terminal, stay with DeepAgents alone. If you want to run it next to other agents with proposals and parallel sessions, add HagiCode. One practical note: the CLI was renamed, so check the setup steps below before connecting it.
LangChain's open-source Deep Agents harness; the terminal coding agent is Deep Agents Code (dcode)
Workbench that runs agent CLIs, DeepAgents included, as executors
Terminal UI with conversation resume, plus headless mode for scripts and CI; Python 3.12+
Desktop app (Windows, macOS, Linux), Docker Compose container, Microsoft Store edition
MIT-licensed and free; you pay your model provider; LangSmith tracing is a separate optional service
Desktop and Container free; Store edition free with optional one-time TurboEngine purchase
OpenAI, Anthropic and Gemini by default; extras for NVIDIA, Ollama and others; any tool-calling model
Hosts no models; passes your chosen model to the CLI; optional OmniRoute routing
Sub-agents with isolated context inside a session; sessions can be resumed
Several sessions at once, each a named Agent with visible status
The agent plans and manages its own context; no proposal lifecycle
OpenSpec proposals: propose, validate, apply, archive (3 concurrent free, 32 with TurboEngine)
Shell access with human-in-the-loop approval; no dedicated commit workflow documented
AI-composed commit messages with the repository locked during the commit; Git and GitHub integration
curl -LsSf https://langch.in/dcode | bash, then dcode
Desktop install, plus the CLI reachable under the command HagiCode launches
Developers who want an extensible LangChain-based harness, sandboxes and tracing
Coding work across repositories with parallel sessions and traceable specs
Deep Agents bundles planning, a filesystem over local, sandboxed or remote backends, context management that summarises long threads and offloads tool output to disk, and sub-agents with isolated context windows. The defaults are tuned for multi-step work rather than single completions.
Deep Agents Code can run code in remote sandboxes such as LangSmith, AgentCore, Daytona, Modal and Runloop. Its README is direct that by default it trusts the directory you run it in and reads project files before any approval prompt, so untrusted repositories belong in a sandbox.
Any model that supports tool calling works, from frontier APIs to Ollama, vLLM or llama.cpp. You can add skills as custom slash commands, persistent memory and MCP tools. The same harness is available as the deepagents Python SDK, and dcode --acp serves it to ACP clients.
Releases land often: deepagents 0.7.26 and deepagents-code 0.1.85 in October 2026. Deep Agents Code was forked from the older deepagents-cli at v0.1.0 in May 2026, and the deepagents-cli package on PyPI is now marked deprecated and repurposed, so use the new name when you search or install.
dcode features the day they ship.HagiCode is a workbench on top of the CLI. It launches the DeepAgents with --acp in the repository you open, so your own install and provider keys do the work. HagiCode adds no model billing.
Nothing is locked in: the CLI remains a normal install you can run with dcode at any time.
curl -LsSf https://langch.in/dcode | bash, run dcode once and configure a model provider.deepagents, then falls back to uvx --from deepagents-cli. If neither works, set the DeepAgents binary path in HagiCode to your dcode executable (dcode --acp is the documented ACP entry; this pairing has not been tested end to end).DeepAgents delegates to sub-agents inside one session. HagiCode adds the level above: several DeepAgents sessions across tasks or repositories, each a named Agent with visible status.
Tracing in LangSmith records what the agent did; an OpenSpec proposal records why. The problem statement, agreed approach and resulting changes stay together and can be archived.
HagiCode drafts the commit message from the diff and locks the repository during the commit. Code Server opens the files the agent edited in a browser editor.
Monospecs lists sub-repositories in .hagicode/monospecs.yaml, so cross-repo work reaches the agent as one proposal. Vault holds reference projects and notes you register once, beside the CLI's own memory files.
OmniRoute can centralise model routing across agents. DeepAgents tool calls appear in the same typed view as every other agent's, and optional TokenMonitor shows today's tokens and cost in the status bar.
Yes. Deep Agents and Deep Agents Code are MIT-licensed. You pay your model provider, and LangSmith tracing is a separate optional LangChain service; see the LangChain pricing page for plans.
Run `curl -LsSf https://langch.in/dcode | bash`, then start it with `dcode`. OpenAI, Anthropic and Gemini support is included; other providers come as extras, for example `DEEPAGENTS_CODE_EXTRAS="nvidia,ollama"`.
Deep Agents Code, package `deepagents-code`, was forked from `deepagents-cli` in May 2026 and is the maintained terminal agent. The old `deepagents-cli` package on PyPI is now deprecated.
Yes, over ACP. HagiCode launches `deepagents --acp`, or falls back to `uvx --from deepagents-cli`; with the renamed CLI you may need to set the binary path to `dcode`.
No. The CLI does the work; HagiCode organises sessions, proposals and commits around it, and you can still run dcode directly.
Pick the CLI alone for sandboxes, SDK work and scripts. Add HagiCode when you run several sessions or repositories and want proposals, commits and Agents managed in one place.
Localized AI code assistant, privacy protection, enhanced efficiency
desktopHero.loading