Hosted MCP
Connect your AI coding agent to your hosted routes, activity and Council settings through the Model Context Protocol (MCP). Hosted MCP is included on every plan, including Free. Your agent runs on your machine; cedr.dev provides account and organization tools.
Let Your Agent Set It Up
Section titled “Let Your Agent Set It Up”- Open MCP in the app under Management.
- Select the organization you want the agent to use in the organization switcher.
- Open Set up with browser sign-in (recommended), or Set up with an API key for a client without browser authentication.
- Review the prompt, select Copy prompt, and paste it into your agent’s chat.
- Follow its sign-in or private secret-setting instructions. The agent verifies setup by discovering the tools and reading your organization’s routes.
The prompts include your server URL and selected organization ID, never a key. They tell the agent to preserve existing servers and configuration, use the client’s supported connection settings, and stop if the client cannot connect directly. The hosted connection is named cedr-cloud so it can coexist with a local cedr server.
Connect Manually
Section titled “Connect Manually”Add this URL as a remote HTTP MCP server in your client:
https://api.cedr.dev/v1/mcpAn example configuration follows. The exact format and authentication commands depend on the client; use its supported remote MCP settings.
{ "mcpServers": { "cedr-cloud": { "type": "http", "url": "https://api.cedr.dev/v1/mcp" } }}This is a hosted server. You do not need to install a local MCP server, run cedr mcp, or keep a terminal open for this connection. A client must support remote HTTP MCP and either browser OAuth or custom authorization headers. Local cedr mcp remains a separate connection for local routing and execution planning.
Browser Sign-In
Section titled “Browser Sign-In”After adding the server, start the client’s sign-in flow. It opens cedr.dev in your browser. Sign in, complete your second factor if enabled, and review the application’s requested access before allowing it. The client manages its own tokens.
| Permission | What it allows |
|---|---|
cedr:read |
Read organization information. Required for every tool. |
cedr:write |
Record verdicts on your own retained synced picks and record Council merges, together with read permission. |
offline_access |
Let the client refresh its access without asking you to sign in each time. |
The connection appears under MCP, then Connected Applications. Application names are supplied by the client, not verified identities. Authorization gives the client access to every organization you belong to, with your existing role and the permissions you allowed. Selecting an organization in the app chooses the setup prompt’s default, not a restriction on the authorization.
API Key
Section titled “API Key”Create a personal key on API Keys. Prefer an expiring, read-only key unless you need to record verdicts. In the client’s secret settings, configure:
Authorization: Bearer YOUR_API_KEYReplace the placeholder privately. Use a supported secret store or environment-variable reference; clients do not all expand variables in JSON. Keep literal credentials out of shared configuration. A key’s expiry and IP restrictions still apply.
API-key clients do not appear under Connected Applications. Manage and revoke them on API Keys. See Authentication for key restrictions and rotation.
Organizations and Tools
Section titled “Organizations and Tools”Every tool takes an org argument containing an organization ID, such as org-3f9a1c0e7b2d. Copy it from MCP, then Organizations, or from Settings. Use the ID, not the organization’s display name; it stays the same when the organization is renamed. Access is checked again on every call; losing membership or changing plans changes what the agent can do immediately.
| Tool | Purpose |
|---|---|
route_list |
List default routes, or saved routes when the organization’s plan includes them. |
route_config |
Read the default or saved organization configuration. |
route_recent |
Read recent synced picks and verdicts, without member emails or notes. Defaults to 25 results, at most 50. |
route_stats |
Read picks, verdicts, accepted rate, time to verdict and cascade outcomes per route and model, with trial versions on their own rows, within the organization’s current retention window. Large summaries can indicate truncation. |
route_machines |
Read permitted machine snapshots and their receipt times. Owners and admins can see all members’ snapshots; members see their own. |
route_verdict |
Record a verdict on your own retained synced pick. Requires read and write access. |
council_members |
Read the configured Council roster, quorum and blocking severity, with optional information from your own machine snapshot. |
council_merge |
Merge submitted JSON review reports, preserving dissent and checking quorum, and return the revision brief when a blocked change has a round left. Records the result’s gate and finding counts; reports are not stored and reviewers are not run. Requires read and write access. |
council_stats |
Read Council results per route: passes, blocks, revised changes and those that passed after revision. |
For a first read, ask your agent to call route_list with your organization ID. No organization data changes.
For route_verdict, supply the original event’s full ISO-8601 time with an offset in at, and reuse that same time on retries. Older events do not overwrite newer verdicts. A hosted verdict changes hosted activity; it is not pulled into the local CLI history.
Snapshots show their age and whether they were received within 15 minutes. They are not live readings and do not make live routing picks. Keep Route sync enabled to send current activity from your machines. Hosted MCP does not provide route_pick, run agents, read local credentials, or access your filesystem. Local CLI use continues to work offline.
Plans and Limits
Section titled “Plans and Limits”MCP connectivity and all tools are included on every plan. Saved organization routes require Pro or above; Free can use default routes. The current plan also controls history retention and request limits.
| Best personal plan | Requests a minute per OAuth connection or API key |
|---|---|
| Free | 30 |
| Pro | 120 |
| Team | 300 |
| Business | 600 |
Personal limits follow your best plan across your organizations. Saved routes and route history follow the plan of the organization named in the tool call. The MCP page shows current limits; see Plans and Route activity retention for details.
Each authenticated MCP message consumes request budget, including setup messages and malformed requests. Requests are limited to 1 MiB. Council merge accepts up to 16 reports, each at most 256 KiB and 1,024 findings. Tool results are capped at 2 MiB; narrow the request if a result is too large.
Monitor and Revoke Access
Section titled “Monitor and Revoke Access”Logs records tool calls and errors, including errors returned inside an HTTP 200 response. Metrics counts those failures. Successful setup messages consume budget but do not fill the activity log. Existing usage and error alerts apply when included in your plan.
Browser-authorized requests appear under OAuth connection; API-key requests use the key’s name. Tool arguments, report bodies, notes and credentials are not logged. The MCP page’s last-request times use retained logs and do not indicate that an agent is online.
To stop an OAuth client, open MCP, find it under Connected Applications, and select Revoke. Its existing tokens stop granting access; reconnecting requires authorization again. To stop a key-based client, revoke its key under API Keys.
Troubleshooting
Section titled “Troubleshooting”| What you see | What to do |
|---|---|
| Sign-in does not open or the client cannot discover tools | Confirm the URL ends in /v1/mcp and the connection is remote HTTP. Check that the client supports browser OAuth or custom headers. Reconnect after changing configuration. |
An expired or revoked connection, or 401 |
Reauthenticate in the client. For a key, check its expiry or revoke it and create a replacement. Browser cookies alone do not authorize MCP calls. |
403 or insufficient permission |
Check membership, the org ID, granted scopes, and the key’s IP restrictions. Read-only access cannot record verdicts or Council merges. Saved routes require an eligible organization plan. |
429 |
Honor Retry-After before retrying and reduce bursts. Every MCP message consumes budget. |
413 or a result-too-large error |
Send fewer or smaller reports, or narrow the requested data. |
| A tool reports an error with HTTP 200 | Read the tool’s error message. Transport success does not mean the tool succeeded; Logs and Metrics record the failure. |
| Snapshots are absent or stale | Sync the relevant machine and check its warnings. Owners and admins can inspect other members’ snapshots; members cannot. |
| A Council result is blocked or incomplete | Review its findings and quorum. A blocked or incomplete gate is a valid result, not a connection failure. |
If you need help, include the tool name, error code and time when contacting support@sheernox.com. Never send credentials, report contents or private tool arguments.