Skip to content

MCP Servers

Canopy ships two distinct MCP (Model Context Protocol) servers - one for Host, one for Cortex - each with its own tool set. There is also a third, unrelated concept (“hosted MCP servers”) that reuses the “MCP” name but is a marketplace feature, not a server you connect an MCP client to. This page covers all three, kept clearly separate.

Package canopy-host-mcp (host/mcp/), built on mcp.server.fastmcp.FastMCP. Exposes read/write tools over the Canopy Host API for coding agents and MCP clients (Claude Desktop, Claude Code, etc.).

  • Transport: stdio by default (CANOPY_HOST_MCP_TRANSPORT), or streamable-http/SSE.
  • CANOPY_HOST_API_URL (default https://api.canopy.pm) + CANOPY_HOST_API_KEY - the server calls the Host API on your behalf using this key.
  • Alternatively, connect with OAuth 2.1: Canopy self-hosts an authorization server at auth.canopy.pm (Ory Hydra). A client that supports OAuth/DCR (Claude Code, Claude.ai, ChatGPT, Codex) discovers it automatically and authorizes through your Canopy account - no API key to copy around. Either way you end up with a bearer token scoped to your own account; there’s no third, separate MCP-level credential.
  • Public URL: https://mcp.canopy.pm, health check at https://mcp.canopy.pm/mcp/health (always reachable, reports host_mcp_hosting_feature_enabled, tool_count, and the live tool list).

Example stdio client config:

{
"mcpServers": {
"canopy-host": {
"command": "canopy-host-mcp",
"env": { "CANOPY_HOST_API_KEY": "your-api-key" }
}
}
}

Feature-gated: the whole tool set is gated behind the host.mcp.hosting platform feature. If it resolves disabled, zero tools are registered (the server stays up, tools/list is just empty) - see Platform Features.

The server registers a default core set; CANOPY_HOST_MCP_TOOL_PROFILE adds opt-in groups (ci, compute, pipelines, graph, governance, library, source, staff) or all. A profile controls which tools are listed, never what you may do: the Host API still authorizes every call. The four host_source_* tools are in the source profile and, like the Canopy Source routes behind them, need the host.source platform feature and source:read.

Tool Params What it does
host_list_projects limit?, cursor? List projects, paginated (25/page by default)
host_get_project project_id Look up one project
host_get_environment environment_id Fetch an environment
host_get_environment_variables environment_id List its variables (values masked - a value_preview only)
host_reveal_environment_variable environment_id, name Real plaintext value of ONE named variable (requires confirmation in clients that support it)
host_get_environment_domains environment_id, limit?, cursor? List its domains, cursor paginated
host_get_environment_runtime_logs environment_id, lines?, process?, app_service_id?, previous? Snapshot of the app’s own stdout/stderr; previous reads the last crashed container
host_create_environment project_id, branch?, repository?, build_path?, build_command?, build_type?, template_id?, name?, migrate_command?, image?, build_only?, package_build?, defer_first_build?, native_bundle_artifact_id?, grove_id?, config? Create an environment (image creates one from an OCI image and starts its first deploy; native_bundle_artifact_id with grove_id creates one from an uploaded native bundle on a native Grove)
host_update_environment environment_id, source_scope?, build_path?, build_command?, build_type?, migrate_command?, health_mode?, backup_consistency?, predeploy_backup?, build_only?, package_build? Update build config, how the app is health-checked, how backups treat volumes, and whether a backup runs before an image change
host_list_builds environment_id, page? List builds (page 1 by default)
host_get_build environment_id, build_id Fetch one build
host_trigger_deploy environment_id, no_cache?, no_reuse?, image? Trigger a deploy (image moves an image environment to a new reference)
host_rollback environment_id, build_id Roll back to a build
host_cancel_build environment_id, build_id Cancel a running build
host_deploy_approval environment_id, build_id, action Approve/reject a gated deploy
host_set_variables environment_id, variables, app_service_id? Bulk-set env vars (for one service when app_service_id is given)
host_mark_variable_secret environment_id, name, app_service_id? Hide a variable’s value everywhere without changing it
host_add_domain environment_id, domain, app_service_id? Add a custom domain (pointing at a web service when app_service_id is given); it gets no traffic until verified
host_get_domain_dns environment_id, domain_id DNS records to create, the ownership TXT record and a live DNS check
host_verify_domain environment_id, domain_id Prove ownership of a domain via its _canopy-challenge TXT record
host_list_previews environment_id, include_destroyed? List a source environment’s PR preview environments (read-only)
host_preview_settings environment_id, enabled?, ttl_hours?, max_concurrent?, fork_policy? Read an environment’s preview settings, or change only the ones given
host_environment_protection environment_id, mode? Read who may reach an environment, or set it to public, authenticated or organization
host_get_preview environment_id, preview_id One preview with its workload phase and databases (read-only)
host_destroy_preview environment_id, preview_id Tear a preview down (destructive, idempotent)
host_list_backups environment_id List an environment’s database backups (read-only)
host_get_backup environment_id, backup_id One backup’s status and metadata (read-only)
host_create_backup environment_id Take a manual backup (rate limited; restore is not a tool)
host_get_restore_status environment_id, restore_id Status of a backup restore (read-only)
host_get_environment_database environment_id The environment’s primary database, no connection details (read-only)
host_list_environment_storage environment_id Persistent storage mounts (read-only)
host_list_volumes environment_id Persistent volumes of a Forest environment: mount path, size, status, retention (read-only)
host_create_volume environment_id, name, mount_path, size_gb?, process?, retention_days?, apply? Create a persistent volume; redeploys the current artifact once unless apply is false
host_detach_volume environment_id, volume_id Detach a volume and keep its data for the retention period (permanent deletion is CLI-only)
host_attach_volume environment_id, volume_id Mount a detached volume again
host_list_adoptable_volumes environment_id Detached volumes of the project this environment can mount (read-only)
host_adopt_volume environment_id, volume_id, mount_path, name?, process?, apply? Mount a detached volume with its data intact
host_get_destroy_plan environment_id What deleting an environment would affect and the retain/delete choices (read-only; deleting is CLI-only)
host_check_environment environment_id Check whether an environment can be deployed without deploying it: image, user, port, volumes, replicas, variables, domains, health (read-only)
host_get_environment_state environment_id One page of health: rollout, restarts and why, volumes, routes, last backup, failing crons, problems (read-only)
host_list_environment_jobs environment_id, kind?, limit? Scheduled commands with next run and recent runs, or recorded one-off and release-task runs (read-only)
host_get_environment_config environment_id The .canopy.yml document the environment deploys with (read-only)
host_get_config_file environment_id, kind, service_id?, include_content? Where a service’s .canopy.yml or canopy.pipeline.ts is in its repository and whether the connection may write it
host_list_config_revisions environment_id Recorded configuration revisions, newest first (read-only)
host_set_environment_config environment_id, content, apply? Record a new configuration revision and redeploy with it unless apply is false
host_get_environment_language_settings environment_id, revision?, service_id? The languages detected in an environment’s source and the build settings they unlock (read-only)
host_set_environment_language_settings environment_id, settings, service_id? Set language build settings (true/false, null resets); applies on the next build
host_analyze_environment_languages environment_id, service_id? Queue a fresh language analysis of the current source revision
host_list_app_services environment_id, kind? An environment’s services (web, worker or job), optionally of one kind (read-only)
host_get_blueprint environment_id An environment’s Blueprint: nodes, links, runtime state, warnings (read-only)
host_plan_blueprint environment_id, document?, patch?, revision? Plan a Blueprint change from a full document, a stored revision number or a list of edits. Changes nothing; returns the ordered operations (immediate or next deploy), impacts, errors with fixes, and a plan id. Needs project:blueprint
host_setup_project team_id, setup Create a project (or add environments to one) with services, Blueprint, domains and first deploys as one durable operation
host_get_setup team_id, setup_id? One setup with per-step status, or the team’s recent setups
host_continue_setup team_id, setup_id, abandon? Run the steps of a failed setup that are not done, or give up and remove what it created
host_resume_blueprint_apply environment_id, apply_id, acknowledge_impacts?, deploy? Finish the newest partially applied or failed apply by re-planning the same Blueprint and running what is left
host_apply_blueprint environment_id, plan_id, acknowledge_impacts?, deploy? Apply a reviewed plan. The plan id is required, so review cannot be skipped. Does not deploy unless deploy is true; a plan with destructive impacts needs acknowledge_impacts. A partially applied or failed apply is a normal result
host_list_blueprint_applies environment_id, apply_id?, limit?, before? An environment’s Blueprint applies, newest first, or one apply with per-operation results (read-only)
host_get_blueprint_status project_id A project’s newest revision, your draft, and each environment’s adoption state, applied revision and last apply (read-only)
host_adopt_blueprint project_id, environment_ids, confirm?, message? Derive the project Blueprint from same-shape environments: a draft to review, or with confirm revision 1 published and the environments adopted
host_draft_blueprint_from_environment project_id, environment_id Replace your draft with what an environment is now, to keep changes made outside the Blueprint
host_get_blueprint_drift environment_id Compare an environment with the revision it last applied (read-only)
host_get_template_blueprint template_id A template read as a Blueprint, with its required variables and notes on what the mapping dropped (read-only)
host_save_template_blueprint template_id, document Write a Blueprint document back to a template’s service rows; per-environment overrides, groups and external hosts are refused
host_plan_template_blueprint environment_id, template_id Plan a template into an environment with its repositories attached; changes nothing until the plan is applied
host_finish_template_blueprint environment_id, template_id, values? After the apply, add the template’s service variables, build steps and required variables; safe to repeat
host_save_blueprint_as_template project_id, name, slug, description?, visibility?, revision? Create a template from a published Blueprint revision, every service kept
host_propose_environment_services repository, branch, project_id?, team_id? Propose a repository’s services from its compose file or Dockerfiles; use team_id before any project exists (read-only)
host_propose_blueprint repository, branch, project_id?, team_id?, provider? Propose a Blueprint from a repository’s root: the document, the evidence behind each node and link, what was left out, and the planner’s dry run. Creates nothing; needs environment:create and the host.blueprint feature (read-only)
host_list_blueprint_revisions project_id, revision?, limit?, cursor? A project’s Blueprint revisions, newest first, or one revision with its document (read-only)
host_diff_blueprint_revisions project_id, from_revision, to_revision What changed between two revisions of a project’s Blueprint (read-only)
host_get_blueprint_draft project_id Your draft of a project’s Blueprint: base and latest revision, whether it is stale, the document (read-only)
host_save_blueprint_draft project_id, document, base_revision Replace your draft with a document edited from base_revision (null when the project has none); 409 if a newer revision exists; publishes nothing
host_rebase_blueprint_draft project_id, choices? Move your draft onto the newest revision; conflicts come back with ids to settle
host_publish_blueprint_draft project_id, message? Store your draft as the next revision and remove it; 409 if a newer revision exists; changes no environment
host_discard_blueprint_draft project_id Remove your draft; published revisions are untouched
host_revert_blueprint project_id, to_revision, base_revision?, message? Add a revision equal to an earlier one; history is kept and no environment changes
host_apply_app_services environment_id, services Configure a new environment from reviewed services, all-or-nothing
host_create_app_service environment_id, service Add a service (Forest only, needs the host.environment_services feature); service.source {repository_id} builds it from another attached repository
host_update_app_service environment_id, service_id, changes Change a service’s settings; applies on the next build; changes.source {repository_id} moves it to another attached repository
host_delete_app_service environment_id, service_id Remove a service (not default, not one with domains or volumes)
host_get_app_service_state environment_id, service_id What one service is doing now, with the age of the Grove’s report (read-only)
host_scale_app_service environment_id, service_id, replicas, process? Set the replicas of a service’s process; 0 pauses a worker
host_service_grove_bindings environment_id, service_id, grove_bindings? Read the machines a service may ask Host for, or replace them (at most 8)
host_restart_app_service environment_id, service_id Roll a service’s pods from the image they already run
host_set_app_service_autoscale environment_id, service_id, process, min_replicas, max_replicas, target_cpu_percent Keep a process’s replicas in a range, steered by CPU
host_run_command environment_id, command, service_id?, process?, with_volumes? Run a one-off command in the default service or in one service’s process
host_run_app_service_job environment_id, service_id, command?, with_volumes? Run a job service once from its live release image
host_get_app_service_execution_output environment_id, service_id, execution_id The end of what a finished job execution printed; needs deploy permission (read-only)
host_list_app_service_executions environment_id, service_id A job service’s recent executions with status and exit code (read-only)
host_list_environment_releases environment_id, limit?, before? An environment’s releases with each service’s image; active marks the live one (read-only)
host_list_environment_repositories environment_id PREVIEW: the repositories an environment builds from, primary first (read-only; needs host.multi_repository)
host_attach_environment_repository environment_id, repository, provider, branch? PREVIEW: attach a secondary repository and register its webhook (needs host.multi_repository and project:blueprint)
host_detach_environment_repository environment_id, repository_id, force? PREVIEW: detach a secondary repository and its webhook; refused for the primary or while a service uses it
host_list_clone_identities team_id PREVIEW: a team’s clone identities, with scope and expiry and never a secret (read-only; needs team:manage and host.multi_repository)
host_rollback_to_release environment_id, release_id Roll back to a release that was live (creates a rollback build)
host_list_team_roles team_id A team’s custom roles and their actions (read-only, preview API)
host_get_team_role team_id, role_id One custom role (read-only)
host_get_my_permissions team_id What you can do on the team, per product action: available, or the one thing in the way (read-only)
host_list_alert_rules team_id A team’s alert rules, without webhook URLs (read-only)
host_list_alert_events team_id A team’s recent alert events (read-only)
host_list_team_images team_id, limit? Docker images a team pushed to the registry, with each tag’s digest-pinned ref (read-only)
host_list_api_keys The caller’s API keys, never key material (read-only)
host_start_job team_id, purpose, image_ref, command, timeout_seconds?, memory_mb?, cpus?, egress? Start an ephemeral isolated job and return at once
host_get_job job_id A job’s state, exit code and output once finished (read-only)
host_list_jobs team_id, limit? A team’s recent jobs (read-only)
host_get_job_events job_id, after? A job’s lifecycle events (read-only)
host_cancel_job job_id Cancel a job that has not started running
host_submit_compute_job team_id, steps, grove? Submit a compute job and return at once; poll host_get_compute_job. Steps run in order on a build Grove. Requires compute:submit; GPU use can be restricted by entitlement and budget
host_get_compute_job job_id, artifacts? A compute job’s status and each step’s state, exit code, error and Grove; artifacts adds the output digests of succeeded steps (read-only)
host_list_compute_jobs team_id, status?, limit?, cursor? A team’s compute jobs, newest first, paged with next_cursor (read-only)
host_cancel_compute_job job_id Cancel a running compute job and ask its Grove to stop what it started
host_deploy_and_wait environment_id, poll_interval_seconds?, poll_timeout_seconds? Deploy, then poll to terminal status
host_artifact_lineage team_id, digest? or commit?, include_impact? Trace an artifact’s lineage
host_artifact_passport artifact_id, sections? Fetch an artifact’s passport ({available: false} if not generated yet); optionally limit returned sections
host_release_and_promote project_id, source_environment_id, target_environment_id, version, build_id?, notes?, poll_interval_seconds?, poll_timeout_seconds? Full release → promote → approve → poll chain
host_catalog_search team_id, q?, limit?, cursor? Search the workflow marketplace
host_catalog_install team_id, workflow_id, granted_permissions Install a workflow
host_catalog_run team_id, installation_id Run an installed workflow
host_environment_recommendations environment_id, limit?, cursor? Read-only rightsizing recommendations
host_compliance_status team_id, limit?, cursor? List compliance evidence packages
host_compliance_package team_id, package_id Fetch one package
host_environment_residency environment_id Residency attestations for an environment
host_artifact_residency artifact_id Residency attestations for an artifact
host_list_ci_connections team_id, limit?, cursor? List Canopy Connect connections
host_list_ci_pipelines connection_id, limit?, cursor? List pipelines for a connection
host_list_ci_runs pipeline_id, limit?, cursor? List runs for a pipeline
host_get_ci_run run_id Fetch a run + its jobs
host_explain_ci_failure run_id Failed jobs only, with failure excerpts (deterministic, no model call)
host_import_ci_connection connection_id Import an existing CI connection
host_get_timeline environment_id, cursor?, limit? Unified build + CI-run timeline
host_cost_summary team_id, since?, until? Cost summary
host_cost_anomalies team_id, threshold_pct? Spend anomalies
host_budgets team_id, limit?, cursor? List budgets
host_pipeline_get_graph graph_id, detail? Fetch a pipeline graph
host_pipeline_get_plan graph_id Fetch a pipeline’s execution plan
host_pipeline_get_run_status graph_id Fetch a pipeline run’s status
host_pipeline_diff_graphs graph_id, against_graph_id Diff two pipeline graphs
host_pipeline_simulate environment_id, ref, commit, untrusted? Simulate a pipeline run without applying it
host_pipeline_validate environment_id, ref?, commit?, content?, service_id?, untrusted? Would this pipeline plan? Pass content to check an unsaved edit (preview)
host_pipeline_detect_dead_steps graph_id Find unreachable/dead steps
host_pipeline_suggest_parallel graph_id Suggest steps safe to parallelize
host_graph_node ref Fetch one engineering-graph node (preview)
host_graph_edges ref, depth? Traverse graph edges (depth clamped 0-3) (preview)
host_graph_commit sha Correlate a commit across the graph (preview)
host_graph_kinds - List engineering-graph node kinds (preview)
host_graph_dependencies ref List a node’s dependencies (preview)
host_graph_impact ref Blast-radius/impact analysis for a node (preview)
host_graph_events ref, after_sequence?, page_size? Events associated with a node (preview)
host_graph_release_live_at environment_id, at? What was live on an environment at a point in time (preview)
host_source_repositories environment_id The repositories an environment builds from, the commits deployed and your provider access (preview, source profile)
host_source_read environment_id, repository_id, path, at? A directory listing or a text file up to 32 KB at a revision; larger or non-text files say so instead of returning content (preview, source profile)
host_source_log environment_id, repository_id, at?, path?, cursor? Commits newest first, 30 per page; continue with next_cursor and the returned sha (preview, source profile)
host_source_compare environment_id, base, head?, repository_id?, path?, node? What changed between two snapshots (every repository), two revisions of one repository (files and commits), or one file’s patch with path; never whole patches (preview, source profile)
host_space_search space_id, query, kinds? Search within a Space, Host-only (preview)
host_deploy_space space_id Deploy a Space (preview)
host_get_space_deploy space_id, space_deploy_id Fetch a Space deploy’s status (preview)
host_list_teams - List teams
host_list_native_groves team_id A team’s native Groves (Macs) with state and auto_placement (read-only)
host_update_native_grove team_id, grove_id, auto_placement?, pools?, portable_steps?, image_builds?, secrets_allowed? Opt a native Grove in or out of automatic placement, replace its pool labels, and set which work it accepts: container steps, image builds, steps with secrets (idempotent)
host_pipeline_secrets team_id, action?, name?, value? A team’s pipeline secrets. action list (names only, the default), set (needs name and value) or remove (needs name). A value is never returned; changing needs variables:write
host_promotion_policies team_id, policy_id? A team’s promotion policies, or one with its document (read-only)
host_list_incidents environment_id, incident_id? An environment’s detected incidents, or one with its timeline (read-only)
host_list_operations team_id, operation_id?, status?, environment_id?, limit?, cursor? A team’s operations timeline of builds, deploys and promotions, or one operation (read-only)
host_list_notifications none Your notifications, newest first (read-only)
host_mark_notifications_read notification_id? Mark one notification read, or all when omitted
host_diagnose_native_grove team_id, grove_id A redacted bundle for one native Grove: state, heartbeat and poll age, the last checks and active_jobs, the steps revoking it would fail (read-only)
host_revoke_native_grove team_id, grove_id Revoke a native Grove: it stops taking work and its running steps fail (destructive; needs grove:manage)
host_list_templates limit?, cursor? List templates
host_get_account - Current user
host_get_plan - Current billing plan
host_list_system_information - What needs your attention now: action items such as an expired git connection or an expiring clone identity, and announcements (read-only)
host_audit_search team_id, event_type?, limit?, cursor? Search the audit event ledger
host_list_webhooks environment_id List git provider webhooks for an environment’s repository
host_create_webhook environment_id Register a webhook, if one isn’t already (idempotent)
host_secure_webhook environment_id Generate/rotate the webhook signing secret
host_delete_webhook environment_id, hook_id Delete one provider webhook by id (from host_list_webhooks)
host_observability_status - Unresolved findings from every platform detection sweep (staff-only, read-only)
host_ops_status - One-page platform health: dead-man report, invariant violations, build queue, Grove heartbeats, dead jobs (staff-only, read-only)
host_list_groves - List every enrolled Forest Grove (staff-only)
host_grove_build_jobs grove_id, status?, build_id?, limit?, offset? A Grove’s build-dispatch jobs (staff-only)
host_grove_workloads grove_id A Grove’s raw workloads rows, incl. sync_cursor (staff-only)
host_list_execution_placements subject_kind, subject_id, node_id? Which Grove each node of a build, sandbox run or environment was placed on, and why (staff-only, at most 500 rows)
host_get_accelerator_inventory substrate?, vendor?, include_empty? Each Grove’s accelerators with allocatable and free counts, its placement units and whether the report is stale, as placement sees them (staff-only; behind the execution observability feature)
host_get_native_groves grove_id?, workload_limit? Native Groves: health, machine (sleep, AC power, disk), workloads with per-instance state, and active jobs (staff-only; behind the execution observability feature)
host_list_grove_placements grove_id, subject_kind?, limit?, cursor? The placements recorded on one Grove, newest first, paged with next_cursor (staff-only; behind the execution observability feature)
host_get_accelerator_entitlement team_id A team’s accelerator entitlement: override, plan columns and the effective value (staff-only)
host_get_accelerator_usage team_id A team’s accelerator-seconds consumed this and last UTC month, with the remaining budget (staff-only)
host_admin_set_accelerator_entitlement team_id, reason, max_accelerators_per_run?, seconds_budget?, budget_mode?, unlimited_budget?, clear_fields? Change a team’s accelerator entitlement override; audited, only the fields passed change (staff-only)
host_list_artifact_blobs team_id, type? A team’s content-addressed artifact blobs (up to 50 per call)
host_pipeline_graphs_reap build_id?, environment_id?, commit?, confirm? Delete stale pipeline_graphs row(s) by build, or by environment and commit (staff-only)
host_admin_set_team_feature_opt_in feature_key, team_id, opted_in, reason Opt a team in to or out of a preview platform feature (staff-only)
host_pipeline_images_audit images Is each pinned image digest reachable from the platform registry and every active build Grove’s own registry? (staff-only, read-only)

Deliberately not exposed: reporting a CI run (that’s a CI credential’s job via canopy ci report, not something an agent should do on a user’s behalf) - see host_explain_ci_failure/host_get_ci_run for the read-only equivalent.

Package cortex-mcp (cortex/mcp/), a separate process with its own tool set for Cortex’s knowledge/RAG/training/agent-run surfaces.

  • Transport: stdio, streamable-http, or sse (CORTEX_MCP_TRANSPORT).
  • HTTP transports require authentication at startup. Choose one mode: CORTEX_MCP_REQUIRE_BEARER_AUTH=true with CORTEX_MCP_BEARER_TOKEN for a shared server secret, or CORTEX_MCP_MULTI_TENANT=true so each caller’s own Cortex API key is forwarded and checked by Cortex. These modes are mutually exclusive. The shared-secret mode authenticates access to the MCP server; it does not give each caller a distinct Cortex identity. The multi-tenant mode does. stdio has no HTTP authentication and is intended for a trusted local client.
  • CORTEX_API_URL (default http://localhost:8082) + CORTEX_API_KEY - used as the ambient identity in single-tenant mode. In multi-tenant mode, each caller’s bearer token supplies the identity and ambient organization defaults are ignored.
  • Scoping defaults: CORTEX_MCP_DEFAULT_PRODUCT_INSTANCE_ID, CORTEX_MCP_DEFAULT_TENANT_ID, CORTEX_MCP_DEFAULT_EXTERNAL_USER_ID.
  • Public URL: https://mcp.canopy.pm, health at https://mcp.canopy.pm/mcp/health.

Feature-gated: the platform manifest classifies cortex.mcp as generally available and internal/staff-only. The MCP server hides all tools only when its feature check receives an explicit disabled result. If the internal state is not visible to its configured API key, or Cortex cannot be reached, the check fails open and tools remain listed. Tool listing is discovery, not authorization: Cortex still enforces each API operation’s permissions and feature checks. A deployment operator must stop or protect the MCP service to revoke access when its feature state cannot be resolved.

The Cortex MCP package includes training and agent-run tools. Their presence in tools/list does not grant API permissions; each Cortex API operation still enforces its own authorization and feature checks.

Tool Params What it does
cortex_list_knowledge_bases - List active knowledge spaces
cortex_create_knowledge_base name, slug, scope_type?, product_instance_id?, description? Create a knowledge space
cortex_retrieve query, knowledge_space_id?, knowledge_space_slug?, product_instance_id?, top_k? Hybrid search, returns ranked chunks with source refs
cortex_ask question, knowledge_space_id?, knowledge_space_slug?, product_instance_id?, assistant_id?, external_user_id?, context_type?, context_ref?, top_k?, max_new_tokens?, temperature?, top_p? RAG Q&A with citations
cortex_list_projects name?, external_tenant_id?, product_instance_id? List tenants/projects
cortex_list_documents tenant_id?, tenant_external_id?, scope?, include_archived?, limit? List ingested documents
cortex_get_context - List available knowledge bases (call first in a session)
cortex_ask_codebase question, external_user_id?, top_k? RAG over Cortex’s own ingested codebase (needs scripts/ingest_codebase.py run first)
cortex_remember content, user_id?, agent_id? Store a persistent memory
cortex_recall query, user_id?, agent_id?, top_k? Search persistent memories
cortex_ingest_and_wait content, name, path?, library_path?, product_instance_id?, poll_interval_seconds?, poll_timeout_seconds? Submit text/markdown for ingestion, poll to completion
cortex_training_run_create name, base_model_name, product_instance_id?, knowledge_space_id?, dataset_id?, machine?, pool?, providers?, gpu_vendor? Create a pending training run; it trains on Canopy, on your own machines first and then Canopy’s GPUs unless machine (one machine), pool (a label), providers (own, canopy) or gpu_vendor (nvidia, apple) limit it
cortex_training_run_list - List training runs
cortex_training_run_launch training_run_id Start a pending run
cortex_training_run_cancel training_run_id Cancel a run
cortex_training_run_status training_run_id Status/loss metrics/error
cortex_agent_run_status run_id Read-only status/steps/chat history for an agent run

canopy mcp (CLI) and sdk.mcpHosting (SDK) are not a way to connect to either server above. They manage preview hosted-server metadata for a team. The registry supports server records, version metadata and credential storage, but Canopy does not currently run the uploaded server: start and invoke return an unavailable error because the runtime is not implemented. The invoke route is a REST-shaped interface, not the MCP wire protocol:

canopy mcp list --team <team-id>
canopy mcp show <server_id> --team <team-id>
canopy mcp create <name> --team <team-id> --namespace <id> --slug <slug>

canopy mcp publish, set-credential, start, invoke, and install commands are also registered. Publishing a version and storing a credential write metadata, but start and invoke fail with an unavailable error while the runtime is absent. install prints the REST invocation URL and tool metadata; the endpoint is likewise unavailable for execution. It is not full MCP-protocol (JSON-RPC/SSE) wire compatibility, so a generic MCP client cannot use it for the standard handshake.

This feature is gated behind host.mcp.hosting, which is generally available. Although the API and CLI can create and publish server metadata, runtime start/invocation is unavailable. See Platform Features for the feature gate and Host CLI for executable examples.