Using the board
Multi-runtime support
Run the board on Claude Code, Codex, Cursor, OpenCode, Kimi Code, DeepSeek Harness, ZCode or Grok Build — and give each agent the one that suits its work.
On this page
A runtime is one coding tool plus the settings the board uses to reach it. The tools run on your machine with your own subscriptions or keys.
Set up a runtime
- Install and sign in to the tool by following its official docs: Claude Code, Codex, Cursor, OpenCode, Kimi Code, DeepSeek Harness, ZCode or Grok Build.
- Add it in Configuration → Runtimes, or run
akb agent runtime add. A new board picks the first installed tool that answers. A tool that is installed but signed out shows Logged out with the command that signs it in. - Fill in its settings from the table below.
- Press Test, or run
akb agent test. It sends one short message, so a bad sign-in shows up now instead of on your next card.
Give each agent its runtime
Every agent uses the board's default runtime until you choose another on its page, or run
akb agent bind <agent> <runtime>. For example, plan on one model and build on a cheaper one.
See Runs. A chat
can switch the agent or model for one conversation.
Settings
| Runtime | Settings |
|---|---|
| Claude Code | Provider (subscription, Anthropic API or a compatible gateway), model, reasoning effort |
| Codex | Provider (ChatGPT subscription, OpenAI API or a Responses API gateway), model, reasoning effort |
| Cursor | Model, Cursor key (optional). Put reasoning effort inside the model id, such as claude-opus-4-8[effort=high] |
| OpenCode | Model as provider/model, reasoning effort. Keys come from opencode auth login |
| Kimi Code | Provider (Kimi sign-in, or a custom endpoint in Kimi, Anthropic or OpenAI format), model |
| DeepSeek Harness | Model, DeepSeek key (optional; otherwise the key dsh saved) |
| ZCode | Model, Z.AI Coding Plan key. Board runs need the key; zcode login alone does not work |
| Grok Build | Model, xAI key (optional; otherwise the key grok login saved) |
Leave the model empty to use the tool's default. To use a gateway such as OpenRouter or LiteLLM, choose Claude Code for Anthropic format, Codex for the OpenAI Responses API, or Kimi Code for any of the three.
Decide what a run may touch
Check each runtime's command and arguments before you start work:
- Restricted writes: OpenCode, DeepSeek Harness and Grok Build keep writes inside the project unless your permission settings widen them. Codex does the same when you choose its workspace-write mode.
- Unrestricted: Claude Code, Cursor, Kimi Code and ZCode run tools without approval prompts and can write outside the project.
- Network: every runtime can reach the network.
If you need a hard boundary, run the board in a container.
What each runtime reports
All eight can chat, run cards and resume a failed run. They differ in what they report in
Runs. akb agent list shows these too.
 |  |  |  |  |  |  |  | |
|---|---|---|---|---|---|---|---|---|
| Early-crash resume | ✓ | — | — | ✓ | — | ✓ | — | ✓ |
| Run cost | ✓ | ✓ | — | ✓ | — | ✓ | — | ✓ |
| Token counts | ✓ | ✓ | — | ✓ | ✓ | ✓ | ✓ | ✓ |
| Rate-limit exit | ✓ | — | — | — | — | — | — | — |
| Provider retry | ✓ | ✓ | — | — | — | — | — | — |
| Direct skill call | ✓ | ✓ | — | — | — | — | — | — |
- Early-crash resume: resumes a run that failed in its first seconds. With the others, press Retry.
- Run cost: an estimate, not a bill.
- Rate-limit exit: ends a rate-limited run and frees the card. The others wait until the limit lifts.
- Provider retry: retries after a temporary provider failure.
- Direct skill call: loads the board skill by name. The others are told to load it in the prompt.
A run that stops producing output ends on its own. Set how long to wait in Configuration → General → End a silent run after. When Claude Code is waiting on its own background tasks, the run does not count as silent. In a chat, the line above the message box shows those tasks and can stop them.
Known limits
- Kimi Code: not yet tested on a real card run.
- ZCode: a run that fails before its first turn has no saved session to resume. Press Retry.