mcp-server

<picture> <source media="(prefers-color-scheme: dark)" srcset="https://github.com/octopusdeploy/mcp-server/blob/main/images/OctopusDeploy_Logo_DarkMode.png?raw=true"> <source media="(prefers-color-scheme: light)" srcset="https://github.com/octopusdeploy/mcp-server/blob/main/images/OctopusDeploy_Logo_LightMode.png?raw=true"> <img alt="Octopus Deploy Logo" src="https://github.com/octopusdeploy/mcp-server/blob/main/images/OctopusDeploy_Logo_LightMode.png?raw=true" /> </picture> # Octopus Deploy Official MCP Server [Octopus](https://octopus.com) makes it easy to deliver software to Kubernetes, multi-cloud, on-prem infrastructure, and anywhere else. Automate the release, deployment, and operations of your software and AI workloads with a tool that can handle CD at scale in ways no other tool can. [Model Context Protocol](https://modelcontextprotocol.io/) (MCP) allows the AI assistants you use in your day to day work, like Claude Code, or ChatGPT, to connect to the systems and services you own in a standardized fashion, allowing them to pull information from those systems and services to answer questions and perform tasks. The Octopus MCP Server provides your AI assistant with powerful tools that allow it to inspect, query, and diagnose problems within your Octopus instance, transforming it into your ultimate DevOps wingmate. For a list of supported use-cases and sample prompts, see our [documentation](https://octopus.com/docs/octopus-ai/mcp/use-cases). ### Octopus Server Compatibility Most tools exposed by the MCP Server use stable APIs that have been available from at least version `2021.1` of Octopus Server. Tools that are newer will specify the minimum supported version in the documentation. Alternatively, you can use the command line argument `--list-tools-by-version` to check how specific tools relate to versions of Octopus. ## 🚀 Installation ### Install via Docker Credentials must be supplied via environment variables to avoid exposing them in the host process list (`ps aux` / `/proc/<pid>/cmdline`). The Octopus server URL can still be supplied via the `--server-url` flag. ```bash docker run -i --rm -e OCTOPUS_API_KEY=your-key -e OCTOPUS_SERVER_URL=https://your-octopus.com octopusdeploy/mcp-server ``` Full example configuration (for Claude Desktop, Claude Code, and Cursor): ```json { "mcpServers": { "octopus-deploy": { "type": "stdio", "command": "docker", "args": [ "run", "-i", "--rm", "-e", "OCTOPUS_SERVER_URL", "-e", "OCTOPUS_API_KEY", "octopusdeploy/mcp-server" ], "env": { "OCTOPUS_SERVER_URL": "https://your-octopus.com", "OCTOPUS_API_KEY": "YOUR_API_KEY" } }, } } ``` For Apple Mac users, you might need to add the following arguments in the configuration to force Docker to use the Linux platform: ```json "--platform", "linux/amd64", ``` We are planning to release a native ARM build shortly so that those arguments will not be required anymore. ### Install via Node #### Requirements - Node.js >= v20.0.0 - Octopus Deploy instance that can be accessed by the MCP server via HTTPS - Octopus Deploy API Key or Access Token (see [Authentication](#authentication) below) #### Configuration Full example configuration (for Claude Desktop, Claude Code, and Cursor): **Write tools enabled (default):** ```json { "mcpServers": { "octopusdeploy": { "type": "stdio", "command": "npx", "args": ["-y", "@octopusdeploy/mcp-server"], "env": { "OCTOPUS_SERVER_URL": "https://your-octopus.com", "OCTOPUS_API_KEY": "YOUR_API_KEY" } } } } ``` **Read-only mode (recommended for production):** ```json { "mcpServers": { "octopusdeploy": { "type": "stdio", "command": "npx", "args": ["-y", "@octopusdeploy/mcp-server", "--read-only"], "env": { "OCTOPUS_SERVER_URL": "https://your-octopus.com", "OCTOPUS_API_KEY": "YOUR_API_KEY" } } } } ``` The Octopus MCP Server is typically configured within your AI Client of choice. It is packaged as an npm package and executed via Node's `npx` command. Credentials (API key or access token) must be supplied via environment variables — they are not accepted as command-line arguments to avoid exposing secrets in the process list. The Octopus server URL may be supplied via either the `OCTOPUS_SERVER_URL` environment variable or the `--server-url` flag. ```bash OCTOPUS_API_KEY=API-KEY \ OCTOPUS_SERVER_URL=https://your-octopus.com \ npx -y @octopusdeploy/mcp-server ``` Or with the server URL on the command line: ```bash OCTOPUS_API_KEY=API-KEY \ npx -y @octopusdeploy/mcp-server --server-url https://your-octopus.com ``` ### Authentication The MCP server supports two authentication methods. Both are supplied via environment variables — credentials are not accepted on the command line because flags are visible in the host process list to any local user. #### API Key (recommended for interactive use) API keys are the standard authentication method for Octopus Deploy. You can generate one from your Octopus Deploy user profile. ```bash OCTOPUS_API_KEY=API-XXXXXXXXXXXXXXXXXXXXXXXXXX \ OCTOPUS_SERVER_URL=https://your-octopus.com \ npx -y @octopusdeploy/mcp-server ``` #### Access Token / Bearer Token (automated scenarios only) The server also supports short-lived access tokens (Bearer tokens) as an alternative to API keys. This authentication method is intended **only for automated scenarios** where an external system issues a short-lived token to the MCP server (e.g., CI/CD pipelines, automated orchestration, or machine-to-machine workflows). Do not use long-lived Bearer tokens — use API keys instead for interactive or long-running sessions. ```bash OCTOPUS_ACCESS_TOKEN=your-short-lived-token \ OCTOPUS_SERVER_URL=https://your-octopus.com \ npx -y @octopusdeploy/mcp-server ``` Full example configuration with an access token: ```json { "mcpServers": { "octopusdeploy": { "type": "stdio", "command": "npx", "args": ["-y", "@octopusdeploy/mcp-server"], "env": { "OCTOPUS_SERVER_URL": "https://your-octopus.com", "OCTOPUS_ACCESS_TOKEN": "YOUR_TOKEN" } } } } ``` If both an API key and an access token are provided, the access token takes precedence. The active authentication method is recorded in the log file (configurable with `--log-file`) so operators can confirm which credential is in use. ### Configuration Options The Octopus MCP Server supports several command-line options to customize which tools are available. If you are not sure which tools you require, we recommend running without any additional command-line options and using the provided defaults. #### Toolsets Use the `--toolsets` parameter to enable specific groups of tools: ```bash # Enable all toolsets (default) npx -y @octopusdeploy/mcp-server # Enable only specific toolsets npx -y @octopusdeploy/mcp-server --toolsets projects,deployments # Enable all toolsets explicitly npx -y @octopusdeploy/mcp-server --toolsets all ``` Available toolsets: - **core** - Basic operations (always enabled) - **projects** - Project operations - **deployments** - Deployment operations - **releases** - Release management - **runbooks** - Runbook discovery and execution - **tasks** - Task operations - **tenants** - Multi-tenancy operations - **kubernetes** - Kubernetes operations - **machines** - Deployment target operations - **certificates** - Certificate operations - **accounts** - Account operations - **interruptions** - Manual intervention and approval operations - **featureToggles** - Inspect and adjust customer feature toggles - **context** - Authenticated user and project context (current user, Git branches) #### Read-Only Mode The server runs with write tools enabled by default. Pass `--read-only` to disable all write tools and block POST/PUT/PATCH/DELETE through the `execute` backstop. Most curated tools are already read-only; only a small set perform writes. **Write-enabled tools (always-write):** - `create_release` - Create new releases - `deploy_release` - Deploy releases to environments and tenants - `run_runbook` - Run a runbook against one or more environments (and optional tenants) - `update_feature_toggle` - Adjust per-environment state and rollout percentages on an existing feature toggle **Conditionally-writing tool:** `execute` is a structured REST backstop whose tier (read / write / delete) is determined by the HTTP method passed to it. See the [API Catalog & Backstop](#api-catalog--backstop) section for details. Write tools are gated by an MCP elicitation prompt: clients that support elicitation will be asked to confirm before the call proceeds. Clients without elicitation support must pass `confirm: true` in the tool arguments — otherwise the tool aborts with an error. Set `OCTOPUS_SKIP_ELICITATION=true` to bypass the gate entirely (intended for unattended automation). The server uses a three-tier read/write/delete classification, enforced server-side based on the HTTP method (the agent cannot bypass this by lying about intent): - **read** — always allowed. GET requests through `execute`, plus all `find_*` / `get_*` / `list_*` tools. - **write** — POST/PUT/PATCH through `execute` and the always-write tools above. Blocked when `--read-only` is set. - **delete** — DELETE through `execute`. Requires `--allow-deletes` and is blocked when `--read-only` is set. A small set of catastrophic-delete paths (e.g. `DELETE /api/spaces/{id}`, `DELETE /api/users/{id}`) and API-key endpoints are on a hard sensitive denylist that ignores both flags. ```bash # Default - write tools enabled (POST/PUT/PATCH) npx -y @octopusdeploy/mcp-server # Additionally permit DELETE requests through the execute tool npx -y @octopusdeploy/mcp-server --allow-deletes # Read-only mode - write/delete tools disabled npx -y @octopusdeploy/mcp-server --read-only ``` **Security Note:** Use an API key with appropriate, least-privilege permissions — write operations can create releases and trigger deployments in your Octopus instance. For production, consider passing `--read-only` unless you have a specific, controlled use case for writes. `--allow-deletes` is off by default; only enable it when the agent must issue DELETE requests through `execute`. If you pass `--allow-deletes` together with `--read-only`, the server prints a startup warning to stderr — DELETE requests remain blocked by the read-only gate. #### Complete Examples All examples below assume `OCTOPUS_API_KEY` is set in the environment. The `--server-url` flag is shown for clarity but can also be provided via `OCTOPUS_SERVER_URL`. ```bash # Development setup with only core and project tools npx -y @octopusdeploy/mcp-server --toolsets core,projects --server-url https://your-octopus.com # Production setup with all tools and read-only enforcement npx -y @octopusdeploy/mcp-server --toolsets all --read-only --server-url https://your-octopus.com # Default invocation - all tools and writes enabled npx -y @octopusdeploy/mcp-server --server-url https://your-octopus.com ``` #### Other command line arguments * `--read-only` - Enable read-only mode: disable all curated write tools and block POST/PUT/PATCH/DELETE through `execute`. Writes are enabled by default; this flag turns them off. See [Read-Only Mode](#read-only-mode). * `--allow-deletes` - Permit DELETE requests through the `execute` tool. Ignored (with a startup warning) when `--read-only` is set. Default `false`. * `--log-level <level>` - Minimum log level (info, error) * `--log-file <path>` - Log file path or filename. If not specified, logs are written to console only * `-q, --quiet` - Disable file logging, only log errors to console * `--list-tools-by-version` - List all registered tools by their supported Octopus Server version and exit ## 🔨 Tools ### URL-Based Tools **Quick start**: Paste Octopus URLs directly to investigate issues without manual ID extraction. - `get_deployment_from_url`: Get deployment details from deployment URL (returns taskId for follow-up) - `get_task_from_url`: Get task details and logs from task URL **Deployment investigation workflow:** ``` 1. get_deployment_from_url with deployment URL → Returns deployment context + taskResourceUri + grepTaskLogHint 2a. Fetch the structured activity tree via resources/read (or read_resource) octopus://spaces/{spaceName}/tasks/{taskId}/details 2b. Or call grep_task_log with the taskId to search the raw log without fetching the full body: grep_task_log({ spaceName, taskId, pattern: "error|fail", caseInsensitive: true }) ``` **Task investigation** (direct task URL): ``` get_task_from_url with task URL → Returns task details and logs immediately ``` These tools eliminate manual ID extraction by: - Parsing URLs automatically - Resolving space IDs to space names - Validating ID formats - Providing clear error messages **Example URLs:** - Deployment: `https://your-octopus.com/app#/Spaces-1/projects/my-app/deployments/Deployments-123` - Task: `https://your-octopus.com/app#/Spaces-1/tasks/ServerTasks-456` See [Working with URLs](docs/working-with-urls.md) for detailed workflows, examples, and best practices. ### Core Tools - `list_spaces`: List all spaces in the Octopus Deploy instance - `list_environments`: List all environments in a given space ### API Catalog & Backstop These tools and resources let the agent reach Octopus REST endpoints that don't have a dedicated curated tool, with hard server-side gating between read, write, and delete operations. - `grep_llms_txt`: Search the Octopus API catalog (`octopus://api/llms.txt`) with grep-style semantics (minimum supported Octopus version: `2026.2.3916`). The catalog body is large (typically 300+ KB) — call this rather than reading the resource body directly. Parameters mirror GNU grep (`pattern`, `caseInsensitive`, `invertMatch`, `fixedString`, `beforeContext`, `afterContext`, `maxCount`). Useful for discovering endpoints (`POST /releases`), enumerating delete endpoints (`DELETE `), or finding the body type for a write operation (`Body: Create.*Command`). - `execute`: Structured REST backstop. Reaches any Octopus REST endpoint under `/api`. The HTTP method is the authoritative read/write/delete classifier — never an `isWrite` flag the LLM can set. Method gating is hard-coded server-side: - `GET` is always allowed (subject to the path shape check + sensitive denylist). - `POST`/`PUT`/`PATCH` are blocked when `--read-only` is set; otherwise they require user confirmation via elicitation. - `DELETE` requires `--allow-deletes` (and is blocked when `--read-only` is set) plus a stronger "IRREVERSIBLE" elicitation message. - The sensitive denylist (API-key endpoints, `DELETE /api/spaces/{id}`, `DELETE /api/users/{id}`) is enforced even with both flags on. - The path is required to be `/api` or start with `/api/` — absolute URLs, SDK-relative `~/api/...` paths, and host-relative paths outside `/api` (e.g. `/octopus/portal/...`) are rejected up front, so `execute` stays bounded to the Octopus REST API surface. - **Per-toolset path allowlist applies only when `--toolsets` has been narrowed.** With every toolset enabled (the default, or explicit `--toolsets all`) the allowlist is bypassed and any path under `/api` is reachable subject to the gates above. When `--toolsets` is narrowed the allowlist becomes the kill-switch: paths only resolve if their owning toolset is enabled, so disabling a toolset (e.g. `certificates`) makes its paths unreachable through `execute` even on `GET`. Catalog data is also exposed as MCP Resources: - `octopus://api/llms.txt` — markdown catalog of every Octopus REST endpoint (HTTP method, path, query params, request/response types). Requires Octopus Server `2026.2.3916` or later. 5-minute in-memory cache keyed on the configured server URL. **Prefer `grep_llms_txt` to reading the body directly.** - `octopus://api/capabilities` — JSON describing the running session: server version, enabled toolsets, available tools (with their `minimumOctopusVersion`), and whether `--read-only` / `--allow-deletes` is on. Useful for the agent to discover what's reachable in this session. ### Projects - `list_projects`: List all projects in a given space ### Deployments - `deploy_release`: Deploy a release to environments (supports both tenanted and untenanted deployments) - `list_deployments`: List deployments in a space with optional filtering ### Releases - `create_release`: Create a new release for a project - `find_releases`: Find releases in a space (can get a specific release by ID, or list/filter releases by project) Release detail is also available as an MCP Resource at `octopus://spaces/{spaceName}/releases/{releaseId}` — fetch via `resources/read` (or the `read_resource` backstop tool) to get the full release body, including release notes and selected packages. ### Runbooks - `find_runbooks`: Find runbooks in a project (can get a specific runbook by ID, or list/filter runbooks by partial name). Each summary includes the published snapshot ID, multi-tenancy mode, and environment scope so callers can pick valid targets before running. - `run_runbook`: Run a runbook against one or more environments. Supports tenanted runs (by tenant name or tenant tag), prompted variables, guided failure mode, scheduled run windows, and step or machine inclusion/exclusion. Defaults to the runbook's published snapshot if `runbookSnapshotId` is omitted. The full runbook body (including runtime policy fields) is available as an MCP Resource at `octopus://spaces/{spaceName}/runbooks/{runbookId}`. ### Tasks Task data is primarily exposed as MCP Resources. Use `resources/read` (or the `read_resource` backstop tool) with one of: - `octopus://spaces/{spaceName}/tasks/{taskId}` — lightweight metadata (state, timing, completion flags) - `octopus://spaces/{spaceName}/tasks/{taskId}/details` — full ServerTaskDetails (Progress, ActivityLogs tree, etc.) For log search, use the `grep_task_log` tool rather than a `/log` resource: - `grep_task_log`: Search a task's activity log without fetching the full body. Parameters mirror GNU grep (`pattern`, `caseInsensitive`, `invertMatch`, `fixedString`, `beforeContext`, `afterContext`, `maxCount`). Returns matching lines with 1-indexed `lineNumber`, optional before/after context arrays, and a `totalMatches` count across the whole log. There is intentionally no `/log` resource: activity logs can be multi-megabyte, and an addressable resource would tempt callers to fetch the entire body when grep is almost always the right primitive. ### Tenants - `find_tenants`: Find tenants in a space (can get a specific tenant by ID or list/search tenants with filters) - `get_tenant_variables`: Get tenant variables by type (all, common, or project) - `get_missing_tenant_variables`: Get tenant variables that are missing values ### Kubernetes - `get_kubernetes_live_status`: Get live status of Kubernetes resources for a project and environment (minimum supported version: `2025.3`) ### Machines (Deployment Targets) - `find_deployment_targets`: Find deployment targets in a space (can get a specific target by ID or list/search targets with filters) ### Certificates - `find_certificates`: Find certificates in a space (can get a specific certificate by ID or list/search certificates with filters) ### Accounts - `find_accounts`: Find accounts in a space (can get a specific account by ID or list/search accounts with filters) ### Interruptions - `find_interruptions`: Find pending or historical interruptions (manual interventions, approvals, guided-failure prompts) in a space, optionally filtered by task, project, environment, regarding document, responsibility, or pending state. Returns slim summaries; dereference the `octopus://spaces/{spaceName}/interruptions/{interruptionId}` resource for the full Form definition (control types, Markdown instructions, button options, submitted Form.Values). ### Feature Toggles - `find_feature_toggles`: List customer feature toggles in a project. Each summary includes per-environment state (`isEnabled`, `rolloutPercentage`, `clientRolloutPercentage`) plus a `resourceUri` so "where is X turned on" is answerable from the list response. - `update_feature_toggle`: Adjust an existing toggle. Narrow surface — flip an environment on/off, change rollout percentages, or update the toggle-level description / default state. Internally fetches the current toggle, applies your patches in memory, and PUTs the merged body, so unmentioned environments and unmentioned fields are preserved. Patches that reference an environment not already configured on the toggle are rejected. The full toggle body (description, tenants, segments, minimum versions) is available as an MCP Resource at `octopus://spaces/{spaceName}/projects/{projectId}/featuretoggles/{slug}`. Rollout group bodies are addressable at `octopus://spaces/{spaceName}/projects/{projectId}/rolloutgroups/{rolloutGroupId}` for read-only inspection. **Out of scope (use the Octopus UI):** creating new feature toggles, deleting toggles, renaming or retagging, attaching/detaching rollout groups, tenant targeting, segments, minimum-version filters, and rollout-group / SDK client-identifier management. ### Additional Tools - `get_deployment_process`: Get deployment process by ID for projects or releases - `get_variables`: Get all project variables and library variable set variables for a project (supports config-as-code projects via `gitRef`) - `get_branches`: Get Git branches for a version-controlled project (minimum supported version: `2021.2`) - `get_current_user`: Get information about the current authenticated user ## 🔒 Security Considerations The Octopus MCP Server includes both read and write operations. Important security considerations: ### Read Operations - Can read full deployment logs, which could include production secrets if they were not marked as secrets - Access to sensitive configuration data and variables - Exercise caution when connecting to tools and models you do not fully trust ### Write Operations By default, the following write operations are available: - **Creating releases**: Can create new releases for projects - **Deploying releases**: Can trigger deployments to environments (including production) - **Running runbooks**: Can execute runbooks against environments and tenants - **Updating feature toggles**: Can flip per-environment state and change rollout percentages on existing toggles - **Arbitrary POST/PUT/PATCH via the `execute` backstop**: Bounded to paths under `/api`, with an always-on sensitive denylist. The per-toolset path allowlist applies only when `--toolsets` has been narrowed; with all toolsets enabled (the default) the only path gates are the `/api` boundary and the sensitive denylist. Pass `--read-only` to disable all of the above. DELETE requests through `execute` require an additional `--allow-deletes` flag — a deliberate opt-in for irreversible operations — and remain blocked when `--read-only` is set. **Critical Security Measures:** 1. **Least Privilege**: Use API keys with the minimum permissions needed for your use case 2. **Opt In to Read-Only Mode**: Writes are enabled by default. For production, pass `--read-only` unless you have a specific, controlled use case for write operations. DELETE always requires the additional `--allow-deletes` opt-in. 3. **Method gating is server-side and hard-coded**: The HTTP method passed to `execute` is the authoritative classifier. The agent cannot bypass the gate by misrepresenting what the call does — POST/PUT/PATCH/DELETE requests get tier-specific gating regardless of the prose in the request body. 4. **Toolset filtering doubles as a kill switch**: Narrowing `--toolsets` removes both the disabled toolsets' curated tools and their paths from the `execute` allowlist. (The allowlist is only consulted when toolsets are narrowed; with all toolsets enabled `execute` is bounded by the `/api` shape check and the sensitive denylist instead.) 5. **Prompt Injection Risk**: Running agents in fully automated fashion could make you vulnerable to prompt-injection attacks **Recommendation**: For production environments, pass `--read-only` unless you have a specific, controlled use case for write operations. Leave `--allow-deletes` off unless you specifically need DELETE semantics through `execute`. ## ⚠️ Limitations ### Data Analysis The nature of current AI chat tools and the MCP protocol itself makes it impractical to analyze large amounts of data. Most MCP clients currently do not support chaining tool calls (using the output of one tool as input to the next one) and instead fall back to copying the results token by token, which frequently leads to hallucinations. If you are looking to process historical data from your Octopus instance for analysis purposes, we recommend using the API directly or writing your own MCP client that is capable of processing the tool call results programmatically. ### Performance The MCP Server is technically just a thin layer on top of the existing Octopus Server API. As such it is capable of retrieving large amounts of data (for example, requesting thousands of deployments). Such queries can have a significant effect on your instance's performance. Instruct your models to only retrieve the minimum set of data that it needs (most models are really good at this out of the box). ## 🤝 Contributions Contributions are welcome! :heart: Please read our [Contributing Guide](CONTRIBUTING.md) for information about how to get involved in this project. We are eager to hear how you plan to use Octopus MCP Server and what features you would like to see included in future version. Please use [Issues](https://github.com/OctopusDeploy/mcp-server/issues) to provide feedback, or request features. If you are a current Octopus customer, please report any issues you experience using our MCP server to our [support team](mailto:support@octopus.com). This will ensure you get a timely response within our standard support guarantees. ## 🙋 FAQ ### Do you have plans to release a remote MCP server? We are working on integrating an MCP server directly into Octopus Server. This will open up the door for us to build more complex MCP tools, as well as: * Giving Octopus Administrators more granular control over MCP clients * Natively support OAuth for client authentication * Integrating security scanning tools into the MCP output If this is of interest to you, please register your interest on [our roadmap item](https://roadmap.octopus.com/c/228-remote-mcp-server-ai-). ## License This project is licensed under the terms of Mozilla Public License 2.0 open source license.