Connections

What this is for. Connections (/connections) is the named inventory of external systems (Odoo, GitHub, MCP, …) that agents may use after you approve them. It is not Channels (messaging accounts such as Telegram) and not the Secrets vault (where credentials are stored). Put the system here; put the password or token in Secrets; put the chat bot under Channels.

Route: /connections.
Subtitle: External systems EYAS can use — inventory, health, and agent proposals.

Connections are a named inventory of external systems (Odoo, GitHub, MCP, …). Credentials go into the Secrets vault; agents can propose a connection for human approval instead of scattering config across MCP, skills, and ad-hoc secrets.

Tabs

Tab	Purpose
Connections	Active inventory (connected / error / disabled / unknown)
Catalog	Known system types — pick one to create an instance
Pending	Agent-proposed connections waiting for Approve / Reject

Connections list

Control / field	Meaning
N connections	Count of inventory rows
Add connection	Open create form (or start from Catalog → Use)
Name	Human label for this instance
System	Catalog type (Odoo, GitHub, …)
Status	Pending / Disabled / Connected / Error / Unknown
Adapter	How EYAS talks to it: `native`, `http`, or `mcp`
Last check	Timestamp of last health test
Error	Last test/error message
Source	User / Agent / System — who created it
Test	Run health adapter (e.g. auth probe)
Edit	Update name, config, secrets
Delete	Remove connection (vault secrets pattern remains documented in Secrets)

Empty: No connections yet. Add one from the catalog or approve an agent proposal.

Create / edit form

Field	Meaning
Name	Display name for this instance
System type	Catalog entry (fixed after create for most flows)
Configuration	Non-secret fields (URL, db, org, …) per system type
Secrets	Sensitive fields — stored in the vault as `conn-{id}-{field}`; never shown again after save
Available to all agents	Default scope when shown
Save / Cancel	Persist or discard

Linked shortcuts: MCP Settings, Secrets (when relevant).

Catalog system types

Type	Adapter	Typical use
Odoo	native	ERP / Helpdesk JSON-RPC + ticket tools
GitHub	http	Repos, issues, PRs, releases
GitLab	http	Projects, issues, MRs
Linear	http	Issues / projects API
Notion	http	Pages and databases
Jira	http	Atlassian Cloud issues
Slack (API)	http	Workspace bot tools (chat channel is separate under Communication)
MCP server	mcp	Link inventory row to an MCP server already configured under MCP
Playwright MCP	mcp	Optional Microsoft `@playwright/mcp` (Apache-2.0, npx). Accessibility snapshot + element refs; optional Playwright MCP Bridge extension for a live tab. Agent tools arrive through the MCP bridge as `mcp_playwright_*`. Test runs a fail-closed doctor (Node 18+, npx). Telemetry off. `--no-sandbox` is forbidden. Not the Python browser-use MCP.
Agent Browser	mcp	Optional Vercel `agent-browser` (Apache-2.0 CLI+MCP). `@e1` refs, state save/load, domain allowlist. Install the CLI, then Catalog → Agent Browser (`mcp --tools core,state`). Test is fail-closed (`EYAS_AGENT_BROWSER_BIN` / PATH). Never `chat`, never the daily Chrome profile. Tools: `mcp_agent_browser_*`.
Chrome DevTools MCP	mcp	Optional Google `chrome-devtools-mcp` (Apache-2.0 npx). Coding/debug: console, network, Lighthouse, WebMCP. Not form-filling. Catalog → Chrome DevTools MCP (`--isolated`). Test is fail-closed (Node 18+, npx). Telemetry off. `--autoConnect` and the daily Chrome profile are refused. Tools: `mcp_chrome-devtools_*`. WebMCP tools only if the sidecar lists them.
Custom HTTP	http	Generic REST with bearer/API-key

Catalog intro: Known system types. Pick one to create a connection instance.

Playwright MCP (optional)

Install from Settings → MCP Servers → Catalog → Playwright MCP, then add a Connections row of this type (mcpServerName defaults to playwright) so Test can doctor it.

Tools appear on the agent as mcp_playwright_* after the MCP server connects. EYAS does not vendor the package.
Live tab: Playwright MCP Bridge extension, then --extension instead of --isolated. Never the daily Chrome/Edge profile.
Doctor is fail-closed, like Hyperframes CLI: missing Node 18+ or npx → Test fails with a remedy. Telemetry is off (DO_NOT_TRACK=1).
Never --no-sandbox / PLAYWRIGHT_MCP_NO_SANDBOX. Never the Python browser-use MCP (uvx browser-use --mcp) — it asks for an LLM API key and exposes retry_with_browser_use_agent.

Agent Browser (optional)

Install the published CLI (npm i -g agent-browser then agent-browser install, or set EYAS_AGENT_BROWSER_BIN), then Settings → MCP Servers → Catalog → Agent Browser. Add a Connections row of this type (mcpServerName defaults to agent-browser) so Test can doctor it.

Tools appear as mcp_agent_browser_*. --tools all / debug are rewritten to core,state (those profiles include chat).
Profile is EYAS-owned (data/browser/agent-browser/profile). Never --profile Default or the daily Chrome/Edge profile (Chrome 136+).
Doctor is fail-closed: missing binary, timeout, or ok: false → Test fails with a remedy. Telemetry off. AI Gateway keys stripped.
Never the npm mcp-agent-browser wrapper. Never agent-browser chat.

Chrome DevTools MCP (optional, coding / debug)

Install from Settings → MCP Servers → Catalog → Chrome DevTools MCP, then add a Connections row of this type (mcpServerName defaults to chrome-devtools) so Test can doctor it.

Tools appear as mcp_chrome-devtools_*. This is not the form-filling lane — use native browser_* for that.
Catalog args: --isolated, --no-usage-statistics, --no-performance-crux, --categoryExperimentalWebmcp=true.
WebMCP (list_webmcp_tools / execute_webmcp_tool) only if the sidecar advertises them (Chrome 150+ with --enable-features=WebMCP). Missing tools are not invented.
Doctor is fail-closed: missing Node 18+ or npx → Test fails with a remedy. --autoConnect and --no-sandbox fail Test.
Never the daily Chrome/Edge profile as --user-data-dir.

See Browser Use for the headless browser_* lane, CLI sidecars, and this coding/debug MCP.

Pending proposals

Agents can call tools to propose a connection. You review reason + config on the Pending tab:

Control	Meaning
Reason	Why the agent wants this connection
Approve	Create/activate the connection
Reject	Dismiss the proposal

Empty pending: No pending proposals.

Agent tools

When the connections module is loaded, agents may use:

Tool	Purpose
`connections_list`	List inventory
`connections_catalog`	List catalog types
`connections_test`	Health-check a connection
`connections_propose`	Propose a new connection for approval