I built Shikigami to run Claude Code, Codex and opencode side by side, each on the model and effort I pick. With delegation turned on, they hand work to each other and report back. You decide the flow, and by default nothing gets submitted until you press Enter.
macOS, Linux & Windows · free, no sign-up
Creates a local branch tracking .
A small Symfony service for team absences, regional holiday calendars, and working-day counts.
Import a regional calendar from a holiday feed without touching recorded absences.
bin/console calendar:import-holidays --region=DENo delegations yet
When an agent in this workspace hands work to another agent, the message appears here with its delivery status.
Every message has been sent or discarded.
Sent to terminal means the text reached the agent’s prompt. It does not tell you whether the agent read it or acted on it.
Shikigami picks a new port every time it starts.
When on, a delegated message is sent without anyone pressing Enter, so agents can drive each other with no one checking, and nothing stops two agents handing work back and forth forever. This is not a security setting.
A read-only reference. Press ⌘ ⇧ / anywhere to open it as a quick overlay.
Shikigami is built and run by one person: Igor Nastarowicz, based in the EU (Poland).
No team, no board, no investors. Just a developer who wanted an easier way to run a fleet of coding agents, and built one.
Free for personal and internal company use.
Redistribution, resale, and repackaging are prohibited; get it only from the official channel.
Enables PHP code intelligence for this project: outline, diagnostics, go-to-definition, find-references and completion. When off, no PHP files are scanned or indexed.
Directory holding composer.json and vendor/, relative to the workspace — e.g. app/. Leave empty if they sit at the workspace root. Changing this rebuilds the index.
Enables Python code intelligence with diagnostics, completion, hover, navigation, and signatures.
Shows the rem equivalent next to font-size declarations written in pixels, calculated against a 16px root. Applies to CSS, SCSS and LESS files.
The loop, the tools, the terminal, the permissions and the context. The model is the engine, and the harness is what lets it drive. Most agent harnesses wrap one model and let it spawn copies of itself as subagents that report back to it.
Shikigami doesn’t replace any of that. Claude Code and Codex keep their own harnesses, subagents included. What I built sits one level up. It lets those agents, from different vendors, see each other and pass work around, and every one of them stays in its terminal that you can read and type into.
I checked each of these against its own docs. They’re good tools, and some of them solve a different problem. The difference I care about is the shape: who can hand work to whom, and who approves communication exchange.
| Aspect | Claude Code agent teams | DeepSeek Harness | Conductor | Codex subagents | Shikigami |
|---|---|---|---|---|---|
| Shape | A fixed lead plus teammates that can message each other | A parent agent delegates to child subagents | Independent agents in parallel workspaces source for Conductor: Shape | A parent spawns subagents and collects their results | Peers: any live agent can delegate to any other |
| Vendors in one setup | Claude Code only | DeepSeek and other providers; Claude Code and Codex as optional child backends | Claude Code, Codex, Cursor, OpenCode, one per chat tab source for Conductor: Vendors in one setup | Codex only | Claude Code, Codex and opencode side by side |
| Agent-to-agent delegation | Yes, between Claude Code sessions | Yes, parent to child; children return a final answer | No; a one-way plan hand-off source for Conductor: Agent-to-agent delegation | Yes, parent to child, Codex only | Yes, in any direction and across vendors, with report back (opencode receives only) |
| Who starts delegated work | Starts automatically; the lead approves teammate plans | An automatic tool call; children run unattended | — | Automatic; subagents inherit the approval mode | You. Messages wait in the prompt for Enter unless you turn on auto-submit |
| Isolation | Teammates share the lead’s working directory | Children share the parent’s workspace | A git worktree and branch per workspace source for Conductor: Isolation | Inherits the parent’s sandbox; the app can use worktrees source for Codex subagents: Isolation | A git worktree or the project root, per agent |
| Where it runs | The Claude Code CLI | Desktop (macOS, Windows), web UI, CLI source for DeepSeek Harness: Where it runs | macOS app; paid cloud workspaces source for Conductor: Where it runs | Codex app, CLI, IDE extension, cloud source for Codex subagents: Where it runs | Desktop app on macOS, Linux and Windows; local only |
| Status & price | Experimental, off by default; paid Claude plan or API | Developer preview; MIT; bring your own keys | Generally available; free tier and paid plans source for Conductor: Status & price | On by default; needs a ChatGPT plan source for Codex subagents: Status & price | Free, closed source |
Save an engine, model, effort and extra flags as a named preset. Quick start then turns “Codex, high effort, in a worktree” into one click.
A library of your own instructions, in groups. Drag one onto an agent and it lands in the prompt unsent, so you can adjust it first.
Agents find their live peers and hand work over as a Task, a Rework or a Blocked request. When they’re done, they report back to whoever asked.
Here’s a flow you can start from:
It’s a recipe, not a pipeline engine. Swap any step’s model, or drop the steps you don’t need.
No single model is the best at every step. I let Codex write the code and have Claude check it, because a reviewer from another vendor doesn’t share the author’s habits. Each agent gets its own engine, model and effort, so the expensive model only runs where it earns its cost.
| Engine | Models | Effort | Delegation |
|---|---|---|---|
| Claude Code | Opus, Fable, Sonnet, Haiku | medium, high, xhigh | sends and receives |
| Codex | GPT-6 Astra, GPT-5.6 Sol / Terra / Luna, GPT-5.5, GPT-5.4 mini | medium | sends and receives |
| opencode | its own catalog, including local Ollama models | — | receives only |
Every delegated message gets its own entry in the Delegations panel. It reaches the receiving agent wrapped in a header that says who sent it, what it’s about, and exactly where it ends. The header tells the agent the message isn’t consent and can’t answer a permission prompt. By default it waits in the prompt until you press Enter.
Delegation is off until you turn it on, and it runs through a local server on 127.0.0.1. The panel marks each entry “Sent to terminal”, “Not delivered” or “Failed”. It never calls a message done just because it was typed in.
Four agents burn through a plan faster than one, so I wanted the limits in one place. The Usage panel shows Claude and Codex plan limits (the 5-hour and weekly windows) with their reset times, plus a context meter for every running agent in every workspace. It reads the files the CLIs already write on disk. Tokens only, no network calls.
Shikigami doesn’t run pipelines for you, and it has no built-in roles. You write the flow and decide who does what. That’s deliberate: I’d rather see every hand-off than trust a runner I can’t watch.
A few boundaries to keep in mind. Agents only reach peers in the same workspace. There’s no “completed” state, because a delivered message was typed in, which isn’t the same as done. opencode can receive delegated work but can’t send it. Usage tracking covers Claude and Codex only.
macOS, Linux & Windows · free, no sign-up
Docs for delegation, prompts and presets are next on my list.