OpenCode Agent
OpenCode Agent puts OpenCode, the open-source coding agent, in your flow with your OpenCode Go subscription, OpenCode Zen credit or any OpenRouter model. It works on a Git repository or an empty folder, in an isolated workspace for your conversation only: it reads the code, changes it, runs commands and tests, and returns its answer and the list of files it changed.
It works as a normal node (an instruction in, the answer out) and as a tool of an Agent or a Voice Agent, so you can ask for changes in your code by voice or chat.
When to use it
- You want an agent to fix a bug, add a small feature, write tests or review a change in a repository.
- You want to program by voice: the voice agent passes on what you say and tells you what was done.
- You want to automate code work from a flow: for example, every new issue produces a branch with a proposed fix.
How it works
- Each conversation (a playground chat, a call, an API session) has its own isolated container with its files. Nobody else sees it.
- Later messages of the same conversation continue the work: the agent remembers what it did and the files are still there.
- The workspace for Run Instruction stops by itself after 15 minutes without use and is kept for 7 days; the next message starts it again as it was. List Sessions, Delete Session and Get Changes run on the fly in a short-lived container destroyed immediately on exit without counting against simultaneous container limits.
- Only Instruction (left) and Tools (bottom) appear on the collapsed node; configure the rest from the side panel or toggle their visibility if you need to wire them.
- The API key reaches only the agent’s process and the Git token only the platform’s own Git commands: the agent never sees the token.
Configuration
| Field | What it does |
|---|---|
| Provider | OpenCode Go (default): your monthly OpenCode Go subscription. OpenCode Zen: pay-as-you-go credit on opencode.ai (includes Claude, GPT and Gemini models). OpenRouter: with an OpenRouter key. |
| OpenCode API Key | With Go or Zen: filled in by the OpenCode credential (the key of your opencode.ai account). |
| OpenRouter API Key | Only with OpenRouter: your OpenRouter key. |
| Operation (tool) | Action to run on the workspace: Run Instruction (default), List Sessions (returns saved sessions with session_name, conversation_id, timestamp and summary), Delete Session (forgets a saved session without touching workspace files), or Get Changes (lists uncommitted changes without spending tokens). |
| Instruction | What to do, in plain words. As a tool, the calling agent writes it. |
| Agent Instructions | Standing rules for every instruction: project conventions, what to avoid, how to answer. |
| Model | A model of the chosen provider, for example kimi-k3 or glm-5.3 on Go, claude-sonnet-5-5 on Zen, anthropic/claude-sonnet-5.5 on OpenRouter. |
| Permissions | What it may do in its workspace: Read only (read and explain), Edit files, or Edit files and run commands (default; needed to run tests). Anything above is refused and listed in the result. |
| Repository URL | The repository cloned the first time. Empty starts from an empty folder. |
| Workspace Name | Unique name to persist the working folder in your Storage (opencode_data/<name>). Left empty, nothing is persisted. Enter the same name on multiple OpenCode Agent nodes or flows to share files and sessions. |
| Skills | Skills of your platform account (yours and those installed from the marketplace) the agent may use. Press refresh to load the list; they are copied to the agent before each instruction. |
| When Done | Commit and push a new branch pushes each instruction’s changes to a new branch nappai/opencode-<date>. |
Advanced options
| Field | What it does |
|---|---|
| Branch | The branch to clone. Empty, the repository’s default branch. |
| Git Access Token | Token for a private repository and for pushing. Only the platform’s Git commands use it. |
| Keep Workspace | During the conversation (default) keeps files and memory between messages; Only for this run discards everything at the end. |
| Session Name (tool) | Logical task name or conversation_id of the session to continue or resume. Empty uses the default session; use a distinct name per task to keep contexts separate inside the same persistent workspace. |
| New Session (tool) | When enabled (true), starts a fresh conversation from scratch instead of resuming the current session. |
| Reasoning Effort | How much the model thinks before acting. Higher is slower and costs more. |
| Output Schema | An optional JSON schema: the agent ends with an answer in that shape, in structured of the result. |
| Time Limit (s) | One instruction taking longer is stopped. Default 600. |
| API Base URL | Only for a compatible gateway of your own. Empty, the provider’s official address. |
Output
- Response: for Run Instruction, the agent’s final answer; for List Sessions, Delete Session and Get Changes, parseable JSON ready for automations (
{"sessions": [...], "count": N},{"changed_files": [...], "diff_summary": "...", "count": N}, or{"deleted": true, "session_name": "..."}). - Result: structured
Data. For Run Instruction it includeschanged_files,diff_summary,branch,denied_actions,steps,conversation_id,session_name,structured,usageandcost_usd; for List Sessions, Delete Session and Get Changes it exposes the structured payload understructuredand at top level (sessions,changed_files,deleted,count). This is what an agent receives when using the node as a tool.
Example
Programming by voice on a repository:
- Add a Voice Agent and a OpenCode Agent with your OpenCode credential.
- Set your repository in Repository URL and, if it is private, the Git Access Token.
- Turn on the node’s tool mode and connect it to the Voice Agent’s Tools.
- Call and say “fix the failing test in the invoices module”: the voice agent passes it on and tells you what changed.
Security
- Always use your API key: subscription logins (claude.ai, ChatGPT) cannot be used in NappAI.
- The agent works isolated: no access to NappAI’s internal network or to other people’s workspaces.
- Only the chosen provider is used: a model it does not have gives an error; it is never swapped for another model or provider.
- Do not write keys or tokens in the instruction: use the credential and the Git Access Token field.