Skip to main content
Tesslate OpenSail

​
What MCP is

Model Context Protocol (MCP) is an open protocol for exposing tools, resources, and prompts from an external server to an LLM agent. OpenSail uses MCP to wire real third-party systems (Slack, Gmail, Google Drive, Linear, GitHub, Brave Search, Postgres, filesystems, and more) into any agent’s tool registry. From the agent’s perspective, MCP tools look identical to built-in tools. The agent calls slack__post_message or linear__create_issue exactly like it calls write_file. The protocol is open, so nothing in your stack is locked to Tesslate.

Real integrations

Slack, Gmail, Drive, Linear, GitHub, Brave, Postgres, filesystem, and more

Per-user credentials

You install a connector with your credentials, encrypted at rest with Fernet

Per-agent binding

Attach a connector to a specific agent or reuse across many

Open protocol

Standard MCP server; BYO server for internal tools

​
Architecture

OpenSail’s MCP layer has three pieces, all on the orchestrator side.
orchestrator/app/services/mcp/client.py opens a streamable-HTTP connection to an MCP server, initializes the session, and lists its tools, resources, and prompts. Streamable HTTP is the only supported transport (see “Why not stdio” below).

​
Why streamable HTTP only

MCP defines two transports: stdio (child process per user per call) and streamable HTTP (stateless HTTP). OpenSail supports only streamable HTTP. Stdio was designed for single-user desktop apps. In a multi-tenant environment it requires Node.js on every orchestrator pod, spawns a process per call per user, and doesn’t scale horizontally. Streamable HTTP offloads load to the connector provider, works identically across replicas, and adds zero dependencies to the backend container. When adding a new MCP server to the seed catalog, verify the server publishes a streamable-HTTP endpoint. If it’s stdio-only, OpenSail cannot use it until the maintainer ships an HTTP transport.

​
The user journey

1

Browse

Open the Marketplace and filter to MCP Servers. Each card shows the server name, the tools it exposes, the auth type (none, bearer), and any required credential fields.
2

Install

Click Install. The connector lands in your Library as a UserMcpConfig. If the server requires credentials, you are prompted to enter them. Credentials are encrypted at rest with Fernet before being stored.
3

Test

From the Library MCP card, click Test Connection. OpenSail opens a streamable-HTTP session to the server, lists capabilities, and returns tool/resource/prompt counts.
4

Attach to an agent

From the Library MCP card, click Add to Agent and pick from your agent list. This creates an AgentMcpAssignment row linking the connector to the agent.
5

Use it

Start a chat with that agent. The worker bridges the connector’s tools into the agent’s registry before the first model call. The session header shows a purple badge for each attached MCP server.
Per-user limits apply: the mcp_max_servers_per_user setting (default 20) caps how many connectors any one user can install. Enough for most workflows, prevents registry bloat.

​
Real connectors

OpenSail’s marketplace ships or is adding official support for:

Slack

Post messages, read channels, manage threads

Gmail

Read, send, label, draft, thread operations

Google Drive

List, search, read, write, and share files

Linear

Create and update issues, comments, projects, cycles

GitHub

Issues, PRs, files, commits, releases

Brave Search

Web search with privacy-first provider

Filesystem

Scoped filesystem access for local workflows

Postgres

Read-only or read-write SQL access

Context7

Library documentation and code examples
More connectors ship on the marketplace regularly. Any streamable-HTTP MCP server works: if your internal system has an MCP server (or can wrap one), you can install it.

​
Managing connectors

Every connector is stored as a UserMcpConfig row for your user. Key operations:
OperationWhere
InstallMarketplace detail page, click Install
Update credentialsLibrary MCP card, Credentials button
Test connectionLibrary MCP card, Test button
Re-discover capabilitiesLibrary MCP card, Details panel, Refresh
Attach to agentLibrary MCP card, Add to Agent dropdown
Detach from agentAgent detail page, MCP tab, remove
UninstallLibrary MCP card, Uninstall (soft-delete)
Tool schemas are cached in Redis for mcp_tool_cache_ttl seconds (default 300). When you rotate credentials or change enabled capabilities, the cache is automatically invalidated.

​
Attaching to an agent

1

Pick the agent

Open the agent detail page from your Library. Go to the MCP Servers tab.
2

Click Add

Pick from your installed connectors. Attachment is idempotent; re-adding a previously disabled assignment re-enables it.
3

Confirm the bridge

Open a chat with that agent. The session header shows a purple plug badge for each attached MCP. The tool list in the agent’s first turn includes prefixed tool names like slack__post_message.
4

Remove when done

Tools can pile up and dilute the agent’s focus. If a connector isn’t being used, detach it. You can always re-attach later.

​
Authoring your own connector

For internal tools, you can run your own MCP server and install it on OpenSail exactly like a marketplace connector.
1

Implement MCP over streamable HTTP

Use the official MCP SDK (Python, TypeScript, or another language) and expose a streamable-HTTP endpoint. Stdio transport is not supported.
2

Declare auth type

Supported auth: none or bearer. If bearer, your clients pass a token via the Authorization header.
3

Host it

Run it anywhere reachable over HTTPS. You can run it inside an OpenSail project if you want it co-located.
4

Register it

Contact Tesslate to add it to the marketplace, or install it directly via API with a custom UserMcpConfig.

​
OpenSail as an MCP server

OpenSail also exposes itself as an MCP server. External coding agents (Claude Code, Cursor, Codex, or your own) can connect to your OpenSail instance and get sandboxed compute, invoke your agents, and manipulate projects through a standard MCP interface. See the Gateway API section of the platform docs for details.

​
Security

All MCP credentials are encrypted at rest using Fernet symmetric encryption. The encryption key is held server-side in CHANNEL_ENCRYPTION_KEY. Credentials are decrypted only at call time and never logged.
Every MCP install and attachment is scoped to your user. Other users (even on the same team) never see your credentials or your attachments.
MCP tools participate in the standard edit-mode gating. If you’re in Ask Before Edit, dangerous MCP tools still prompt for approval before execution.
Rotate credentials in the Library. Uninstall to soft-delete the config and invalidate the schema cache. Tokens can be revoked at the provider at any time; OpenSail will fail closed on the next call.

​
Troubleshooting

Check that the AgentMcpAssignment exists and is enabled. Verify the chat is using the expected agent (worker logs show MCP context query: agent_id=...). Re-run Test Connection from the Library to force a fresh discovery.
The MCP server is unreachable. Check the configured URL, the server’s uptime, and any network restrictions. Some servers rate-limit by user key; verify your credential is still valid.
This error means the connector’s config specified stdio. Only streamable HTTP is supported. If the connector’s maintainer doesn’t ship an HTTP endpoint, OpenSail cannot install it.
Default MCP tool timeout is 30 seconds (mcp_tool_timeout). Some tools (large Drive searches, complex Postgres queries) need longer. The setting is orchestrator-wide; contact your admin if you’re self-hosting.
Schemas are cached in Redis for 5 minutes by default. If a connector added new tools and the agent isn’t seeing them, run Re-discover from the Library card. This invalidates the cache and fetches fresh schemas.

​
Next steps

Customizing Agents

Bind connectors to custom agents

Skills

Pair connectors with skill playbooks for end-to-end workflows

API Keys

Manage external API keys for model BYOK and agent invocation

Using Agents

See MCP tool calls in the chat transcript