Synapse vs Claude Code

At a glance

Claude Code facts are from the Claude Code docs on parallel agents and linked pages, checked on 16 October 2026. The newest version in the changelog was v2.1.301.

What Claude Code's multi-agent features are

Over 2026 Anthropic added many ways to run more than one Claude at once:

• Subagents run with their own context window and return a summary. Up to 20 can run at once by default.
• Agent teams have a lead and teammates with a shared task list, file-locked task claiming and a mailbox. They are experimental and off by default. The docs say teams use about 7x more tokens than a standard session when teammates run in plan mode.
• Agent view (claude agents, research preview) shows all background sessions on one screen, each in its own git worktree.
• Desktop, web and cloud sessions keep running after you close the app.

If all your agents are Claude, this first-party toolkit covers a lot.

What Synapse is

Synapse puts every live coding agent on your machine on one canvas, grouped by git repo. It finds Claude sessions from Claude Code's own session registry and every other CLI by scanning processes.

Drag from agent A to agent B and B gets A's context. On each prompt, a hook injects what A said since last time, capped at 8,000 characters per link. Linked agents can also message each other through MCP tools, a CLI and a skill.

How Claude Code agents talk to each other

This is the part people search for, so here it is precisely.

Subagents report back to whoever started them.

Agent teams let teammates message each other by name through a JSON mailbox. Teammates do not inherit the lead's conversation. They get CLAUDE.md, MCP servers, skills and the spawn prompt.

Cross-session messaging (v2.1.224 and later, on by default) lets Claude list your other Claude Code sessions with /list-agents or /peers and send text with SendMessage. You can @session-name in a prompt and subscribe to idle notifications. It reaches local, cloud and Remote Control sessions. Inbound messages cannot approve prompts or change config, and you can set sessions to accept, hold or refuse them.

Two limits matter for this comparison. In the docs' words: "A message is a piece of text one Claude writes to another, never the sender's conversation history or files." And every participant must be a Claude Code session. A Codex or Gemini agent cannot show up in /list-agents.

Synapse works differently on both points:

• A link carries recent conversation, not just a note. If your Claude agent is linked from a Codex agent, it sees what Codex actually said since its last prompt.
• Links work across vendors: Claude Code and Codex both ways, Gemini CLI reading Claude, and messages to OpenCode, Cursor Agent, Amp and Copilot CLI.
• Messages are pulled, not pushed. An idle agent does not wake up, so Claude's idle notifications are more responsive here.

OpenAI's official Codex plugin for Claude Code adds /codex:review and /codex:rescue to hand a task to Codex as a subagent. That is one-way delegation, not a link between two running agents.

Forks: /fork, /subtask, /branch, /btw and /sfork

Claude Code's forking is native, shipped and richer than Synapse's. Here is the current set:

• /subtask starts a fork that inherits the whole conversation so far and returns only its final result to the main context. It was called /fork until v2.1.212.
• /fork, with agent view on, copies the whole session into a new background session that runs alongside it.
• /branch copies the conversation and switches you into the copy. The --fork-session CLI flag does the same from the command line.
• /btw asks a side question with the session's context without adding it to the main thread. In the Desktop app it is Cmd+;.

Synapse's /sfork is in development in its desktop app, for Claude Code and Codex. It calls the same native command (claude --resume <id> --fork-session) and stacks the fork as a card on its parent, so the side question never enters the main context. The lineage shows on the canvas as a dashed line, and the child reports back only if you ask.

Honestly: if you only use Claude Code, /btw and /subtask already solve side questions well. /sfork adds the same gesture for Codex and a visual fork tree.

Accounts, privacy and lock-in

Claude Code needs a paid Claude plan (Pro from $20 a month) or an API key. The free plan does not include it. Same-machine messaging goes over a local socket and, per the docs, "never through Anthropic servers". Cross-machine and cloud messaging go through Anthropic and need claude.ai sign-in plus Remote Control.

The bigger lock-in point is the one the docs state plainly: "In every approach the workers are Claude sessions. To involve a different tool, expose it to Claude as an MCP server."

Synapse has no account. Links and messages live in ~/.synapse, and it makes no network calls. Shared context is labelled as coming from peers, not the user.

Where Claude Code is ahead

• It is native. Task lists with dependencies, file locking, idle notifications and permission rules for inbound messages. Synapse has none of that.
• Forks are shipped. /subtask, /fork, /branch and /btw work today. Synapse's forks are in development.
• Cross-machine works now. Remote Control and cloud sessions can message each other. Synapse multiplayer is not shipped.
• More surfaces. Agent view, Desktop, web and cloud sessions that keep going with the laptop closed.
• Maturity. Claude Code has a huge user base. Synapse is new, built in October 2026, macOS first.

Using Synapse with Claude Code

1. Clone the repo and run node install.mjs && npm start. The installer adds an MCP server, a prompt hook and a /synapse skill, backing up each config first.
2. Restart running Claude Code sessions so they pick up the hook.
3. Open the canvas at localhost:4747. Claude, Codex and Gemini agents appear as nodes.
4. Drag a link from a Codex agent to your Claude Code lead. The lead now gets Codex's recent conversation on every prompt.
5. Keep agent teams and subagents for Claude-internal work, and Synapse for the edges to other vendors.

Who should pick which

• Stay with Claude Code's built-ins if every agent you run is Claude. Teams, subagents, /subtask and cross-session messaging are more complete than Synapse for that.
• Add Synapse if you mix Claude Code with Codex, Gemini CLI or others and want them to see each other's work.
• Use both for mixed fleets.

FAQ

Can Claude Code agent teams include Codex or Gemini agents?

No. Claude Code's docs say every worker in every approach is a Claude session, and other tools can only be exposed to Claude as MCP servers. Synapse can link a Claude Code agent to a running Codex or Gemini CLI agent on the same machine.

Does Claude Code cross-session messaging share conversation history?

No. Claude Code's cross-session messages are plain text one Claude writes to another, never the sender's conversation history or files. Synapse links inject the linked agent's recent conversation, up to 8,000 characters per link, on each prompt.

What is the difference between /fork, /branch and /subtask in Claude Code?

In current Claude Code, /subtask starts a fork that inherits the conversation and returns only its result, /fork with agent view on copies the session into a new background session, and /branch copies the conversation and switches you into the copy. /btw asks a side question without adding it to the main thread.

How do I make Claude Code talk to Codex?

Install Synapse, restart your agents, and link the Claude Code and Codex nodes on the canvas both ways. Each then sees the other's recent conversation and they can message each other. OpenAI's Codex plugin for Claude Code delegates tasks one way instead.

Is Synapse's /sfork the same as Claude Code's /fork?

They use the same native fork underneath, but Synapse's /sfork is still in development. It stacks forks on their parent, draws the tree on a canvas and also works for Codex. Claude Code's forking is shipped.

Does Synapse need a Claude account?

No. Synapse has no account and makes no network calls. Claude Code still uses your existing plan or API key.