SideButton Claude connector: dispatch agents from any chat
The Claude connector turns any Claude conversation into a dispatch console for your SideButton agent fleet. Ask Claude to run an SE or QA workflow on a Jira ticket, and an agent VM picks it up, does the work, and reports back.
What the Claude connector does
The SideButton cloud MCP behind the Claude connector adds 26 tools to Claude, in four groups:
- Fleet & dispatch — see your agents and their health, and send a workflow (e.g.
se-work,qa-validate) to one agent or queue it for several tickets. - Tasks & runs — review the task pool, explode one ticket's runs and transcripts, and run a whole playbook on a ticket.
- Skill pack — read and edit your account's knowledge pack, and sync it.
- Playbooks, roles & routing — read and change the playbooks, roles, workflows and issue-type routing your runs use.
Agents run on Linux VMs you provision through your SideButton account. Each agent has its own Claude Code session, browser, and installed knowledge packs. The cloud MCP just gives Claude a remote control for the fleet — the actual work happens on the agent VM.
Requirements
- A SideButton cloud account. Sign in at sidebutton.com/portal.
- At least one provisioned agent VM visible in your portal's Agents tab.
- Workflows configured for your agents (SE, QA, PM — or custom).
- A Claude client that supports custom connectors (Claude.ai, Claude Code, Claude Desktop).
Set up the connector
1. Add the connector in Claude
In Claude settings, go to Connectors → Add custom connector and enter:
https://sidebutton.com/mcp 2. Sign in
Claude will open an OAuth 2.1 flow. You will be redirected to SideButton, then to our identity provider (Auth0) to sign in with your SideButton account. Claude then holds its own token for your account, which renews itself and which you can revoke on its own.
3. Verify
In your next Claude message, try:
List my SideButton agents and show their status.
Claude will call list_agents and report back. If you see your fleet,
you're ready to dispatch work.
Tools reference
| Tool | Type | What it does |
|---|---|---|
list_agents | read-only | List Agents |
list_workflows | read-only | List Workflows |
list_playbooks | read-only | List Playbooks |
get_playbook | read-only | Get Playbook |
get_issue_type_routing | read-only | Get Issue Type Routing |
get_pack | read-only | Get Pack |
manage_playbook | write | Manage Playbook |
manage_role | write | Manage Role |
manage_workflow | write | Manage Workflow |
set_issue_type_routing | write | Set Issue Type Routing |
get_agent_status | read-only | Get Agent Status |
dispatch_agent_job | write | Dispatch Agent Job |
queue_jobs | write | Queue Jobs for Tickets |
run_workflow | write | Run Workflow Now |
queue_workflow | write | Queue Workflow |
get_task | read-only | Get Task |
review_tasks | read-only | Review Tasks |
work_on_task | write | Work on Task |
get_run_transcript | read-only | Get Run Transcript |
list_pack_files | read-only | List Pack Files |
read_pack_file | read-only | Read Pack File |
write_pack_file | write | Write Pack File |
scaffold_pack_node | write | Scaffold Pack Node |
delete_pack_node | write | Delete Pack Node |
sync_pack | write | Sync Pack |
publish_task_review | write | Publish Task Review |
Every tool declares its type to Claude (readOnlyHint or
destructiveHint) so Claude can ask for confirmation before dispatching
work. Tool names are under 64 characters; no tool mixes read and write operations.
The tools that change your pack, playbooks, roles, workflows or routing need the admin role.
Example prompts
- Dispatch a single job:
Run se-work for SCRUM-142 on agent 5 with high effort. - Queue a batch:
QA-validate SCRUM-120, 121, and 122. Use max effort. - Check fleet health before dispatching:
Which of my agents is free right now? Send the next ticket there. - Run an ad-hoc workflow:
Run nightly-backup on agent-11 now.
Security & privacy
- OAuth 2.1 — authorization code flow with PKCE (S256), Dynamic Client Registration per RFC 7591.
- Metadata — RFC 8414 and RFC 9728 discovery documents at
/.well-known/oauth-authorization-serverand/.well-known/oauth-protected-resource. - Origin validation — MCP endpoints reject requests from unknown browser origins (DNS-rebinding protection).
- Scope — the connector only sees your account. No cross-tenant access.
- Data — details in our Privacy Policy, section 2.5 (Cloud Services).
- Revocation — remove the connector from Claude settings, or revoke it in the portal under Settings → Assistant → Connect an MCP client, where every client you connected is listed with its last use.
Troubleshooting
HTTP 401 on tool calls
The connector was revoked, or its sign-in lapsed after a month unused. Reconnect it in Claude.
HTTP 403 "forbidden_origin"
Your client is sending an Origin header we don't recognize. Allowed origins are
claude.ai, claude.com, this site's own
(sidebutton.com), and loopback. Non-browser clients (Claude Code, curl) usually send no Origin
header and pass through.
Agent is offline
Open the portal's Agents page to see agent health. An offline agent cannot be dispatched to — try another agent, or check the VM from the portal.
"Workflow not found"
Workflow IDs are case-sensitive and must match what your account has installed.
Call list_workflows first to see the exact names.
More resources
- Open the portal — provision agents, install workflows, manage tokens.
- Privacy Policy — what cloud data we store and why.
- Terms of Service.
- GitHub Issues — support and bug reports.