Coding with agents
Point Claude Code, Cursor and other coding agents at these docs, and connect them to your pod.
A coding agent can connect to two different things. They are separate, and you can give it both.
| What the agent gets | Address | |
|---|---|---|
| These docs | The documentation, to read and search while it writes code | https://docs.synap.live/llms.txt, any page as .md, the docs MCP server at /api/mcp |
| Your pod | Your data, under the agent's own key and governance | Your pod's /mcp, connected with @synap-core/cli |
Give it the docs when it writes code against Synap (an app on the SDK, a script on the Hub REST API, an MCP client). Connect it to your pod when it should work with your data.
Give an agent these docs
Start from llms.txt
/llms.txt is the entry point written for agents: the sections, the key pages,
the interfaces to prefer, and the package names that are retired and must not be used.
/llms-full.txt is every page in one file.
To make an agent use them in a repository, say so in its instruction file (CLAUDE.md, AGENTS.md or a Cursor
rule):
Synap documentation: https://docs.synap.live/llms.txt
Read a page as Markdown by adding .md to its URL.Every page as Markdown
Each page has a Markdown version, built from the same source as the HTML:
curl https://docs.synap.live/docs/build/agents/mcp.md # add .md to the URL
curl -H 'Accept: text/markdown' https://docs.synap.live/docs/build/agents/mcp # or ask for MarkdownOn every page, Copy Markdown copies it, and Open sends it to ChatGPT, Claude or Cursor, or shows it on GitHub.
The docs MCP server
https://docs.synap.live/api/mcp is an MCP server (Streamable HTTP, no sign-in) with three tools:
| Tool | Does |
|---|---|
list_pages | The index of every page, with its URL |
get_page | One page as Markdown, by its URL (for example /docs/build/sdk) |
search | Full-text search over the docs |
Claude Code:
claude mcp add --transport http synap-docs https://docs.synap.live/api/mcpCursor, in ~/.cursor/mcp.json or a project's .cursor/mcp.json:
{
"mcpServers": {
"synap-docs": { "url": "https://docs.synap.live/api/mcp" }
}
}This server only serves the public documentation. It cannot see a pod.
Connect an agent to your pod
Every pod serves its own MCP server at /mcp. The CLI connects an agent to it with an identity and a key of its own,
so its writes go through your pod's governance and you can revoke it separately. See
Connect an agent for every client.
npx @synap-core/cli init # first time: finds or creates your pod and connects the agents it detects
synap connect claude-code # or: codex, cursor, claude-desktop, ...- Claude Code:
synap connect claude-coderegisters the pod's MCP server withclaude mcp add(user scope) and installs the Synap skills in~/.claude/skills/. - Cursor:
synap connect cursorwrites the pod's MCP server into~/.cursor/mcp.json. - claude.ai (web): its connector settings take a URL and sign you in with OAuth, with no key to paste.
synap mcp connect-claudeprints the steps and the Synap Cloud URL,https://api.synap.live/mcp. After you approve, ask Claude to list your pods (list_pods) and connect one (connect_pod).
Check the connection with synap mcp verify and synap whoami.
What the agent sees
- Tool names start with
synap_. A new agent key lists a small entry set:synap_orientis the session briefing (what awaits your review, open work, your identity and projects), andsynap_askanswers questions from your data. The agent loads more tools in groups. MCP tools lists them all. - Its writes become proposals unless a rule you set lets them apply. It can never approve a proposal.
- What it may read and change is set by its key: see API keys.
Both at once
The docs MCP server and your pod's MCP server are separate entries in the client's configuration: for example
synap-docs, and the synap-<pod name> entry that synap connect writes. An agent building on Synap uses the first to look up how, and the second to act on your
data.