Implementing MCP Tools

Implementing MCP tools

Read the full guide at docs/published/handbook/engineering/ai/implementing-mcp-tools.md.

Quick workflow

# 1. Scaffold a starter YAML with all operations disabled.
#    --product discovers endpoints via their x-product attribution.
#    ViewSets in products/<name>/backend/ are auto-attributed via module
#    path. ViewSets elsewhere need
#    @extend_schema(extensions={"x-product": "<product>"}).
pnpm --filter=@posthog/mcp run scaffold-yaml -- --product your_product \
    --output ../../products/your_product/mcp/tools.yaml

# 2. Configure the YAML — enable tools, add descriptions, and annotations for PATCH/POST/PUT
#    (scopes come from the API when omitted)
#    Place in products/<product>/mcp/*.yaml (preferred) or services/mcp/definitions/*.yaml

# 3. Add a HogQL system table in posthog/hogql/database/schema/system.py
#    and a model reference in products/posthog_ai/skills/querying-posthog-data/references/

# 4. Generate handlers and schemas
hogli build:openapi

# 5. Refresh the tool input schema snapshots (CI unit tests fail on a stale snapshot)
pnpm --filter=@posthog/mcp exec vitest run tests/unit/tool-schema-snapshots.test.ts -u
# A tool behind a new `feature_flag` needs that flag in the test's `featureFlags` map, set to the value that shows the tool:
# true for a plain gate, the variant string for a variant gate, a non-true value for a `disable` gate.

# 6. Only when the YAML uses ui_apps: regenerate the UI apps (CI checks they are current)
pnpm --filter=@posthog/mcp run generate:ui-apps

Before you scaffold: fix the backend first

The codegen pipeline can only generate correct tools if the Django backend exposes correct types. Read the type system guide for the full picture.

Before scaffolding YAML, verify:

1. Serializers have explicit field types and help_text — these flow all the way to Zod .describe() in the generated tool. Missing descriptions = agents guessing at parameters. Use ListField(child=serializers.CharField()) instead of bare ListField(), and @extend_schema_field(PydanticModel) on JSONField subclasses to get typed Zod output (see products/alerts/backend/presentation/views/alert.py for the pattern).
2. Plain ViewSet methods have @extend_schema(request=...) — without it, drf-spectacular can't discover the request body and the generated tool gets z.object({}) (zero parameters). ModelViewSet with a serializer_class is fine; plain ViewSet with manual validation is not.
3. Query parameters use @validated_request or @extend_schema with a query serializer — otherwise boolean and array query params may produce type mismatches in the generated code.

If a generated tool has an empty or wrong schema, the fix is almost always on the Django side, not in the YAML config. For a full audit checklist and before/after examples, use the improving-drf-endpoints skill.

When to add MCP tools

When a product exposes API endpoints that agents should be able to call. MCP tools are atomic capabilities (list, get, create, update, delete) — not workflows.

If you're adding a new endpoint, check whether it should be agent-accessible. If yes, add a YAML definition and generate the tool.

Tool design

Tools should be basic capabilities — atomic CRUD operations and simple actions. Agents compose these primitives into higher-level workflows.

Good: "List feature flags", "Get experiment by ID", "Create a survey". Bad: "Search for session recordings of an experiment" — bundles multiple concerns.

Tool naming constraints

Tool names and feature identifiers are validated at build time and in CI. Violations fail the build.

Tool names

• Format: lowercase kebab-case — only [a-z0-9-], no leading/trailing hyphens
• Length: 52 characters or fewer
• Convention: domain-action, e.g. cohorts-create, dashboard-get, feature-flags-list

Keep action verbs out of compact tool domains

The single-exec prompt builds its compact domain index with ToolDomainExtractor. Large families can split at an intermediate segment. Without action trimming, experiment-freeze-exposure can advertise the redundant domain experiment-freeze instead of experiment.

Whenever you add or rename an action tool, check the TRAILING_ACTIONS set in the same change. If a rendered domain can end in an operation verb that is not already present, add the verb. Cover it in services/mcp/tests/unit/instructions.test.ts. This applies even when the verb is not the final segment of the full tool name.

Add operation verbs such as freeze, publish, or emit. Do not add resource or capability nouns such as config, logs, stats, or schedule merely to make the prompt shorter; those remain useful discovery domains.

Feature identifiers

• Format: lowercase snake*case — only [a-z0-9*], must start with a letter
• Convention: should match the product folder name, e.g. error_tracking, feature_flags

Why 52 characters?

MCP clients enforce different limits on tool names. The 52-char limit is the safe zone that works across all known clients:

With the server name "posthog" (7 chars) plus a separator, tool names must stay at or below 52 characters to fit within Cursor's 60-char combined limit.

CI enforcement

• pnpm --filter=@posthog/mcp lint-tool-names — validates length and pattern for YAML and JSON definitions
• A vitest test validates all runtime TOOL_MAP and GENERATED_TOOL_MAP entries

YAML definitions

YAML files configure which operations are exposed as MCP tools. See existing definitions for patterns:

• products/<product>/mcp/*.yaml — preferred, keeps config close to the code
• services/mcp/definitions/*.yaml — fallback for functionality without a product folder

The build pipeline discovers YAML files from both paths.

Key fields

category: Human readable name
feature: snake_case_name # should match the product folder name (used for runtime filtering)
url_prefix: /path # frontend app route, used for enrich_url links
tools:
  your-tool-name: # kebab-case
    operation: operationId_from_openapi
    enabled: true
    # Optional:
    scopes: # defaults to the scopes the API requires (from the OpenAPI spec)
      - your_product:read
    annotations: # defaults for GET and DELETE; required for PATCH, POST and PUT
      readOnly: true
      destructive: false
      idempotent: true
    title: List things
    description: >
      Human-friendly description for the LLM.
    list: true
    enrich_url: '{id}'
    param_overrides:
      name:
        description: Custom description for the LLM
    response: # filter response fields (applied per-item on list endpoints)
      include: [id, key, name] # keep only these fields (dot-path wildcards supported)
      exclude: [filters.groups.*.properties] # remove these fields
      # include and exclude are mutually exclusive
      selectable: true # add optional `fields` param so the agent picks a subset of `include` per call
      # (constrained to the allowlist); omit `fields` to return the full set. Requires `include`.
      strip_nulls: true # remove keys whose value is `null`, applied after include/exclude
      # Use it on tools that echo a nested serializer schema, where the unset optional fields
      # dominate the payload. Rejected with `list: true`, where per-row null removal makes the
      # TOON table larger. Use `exclude` to drop the fields on a list tool instead.
    feature_flag: my-flag-key # gate this tool behind a PostHog feature flag
    feature_flag_behavior: enable # 'enable' (default) or 'disable'

When scopes is omitted, the generator uses the scopes the API requires, so the tool cannot drift from the endpoint. Set scopes by hand only when the API computes them per request (the generator fails and says so) or to gate a tool more tightly. A scopes list that misses a scope the API requires fails codegen, with a GitHub annotation on CI. When the API picks the scopes per request, list the action in the viewset's request_dependent_scope_actions. The spec then marks the operation with x-request-dependent-scopes, codegen skips the check for it, and its tools must declare scopes. annotations default to the HTTP method for GET (read-only) and DELETE (destructive). PATCH, POST and PUT vary too much (a PATCH can be a soft delete or non-idempotent), so declare annotations for them.

When a tool needs custom logic around the request, set hooks: <path under src/tools/> and default-export an object with beforeRequest, afterResponse or onError from that module, written export default { onError } satisfies ToolHooks<Params> so a typo fails typecheck (see ToolHooks in src/tools/tool-hooks.ts). Use it to read state before a write or to turn a known error into a result, rather than shadowing the generated tool with a hand-written one.

Unknown keys are rejected at build time (Zod .strict()).