v0.3.0-alpha.1

[0.3.0-alpha.1] - 2026-07-04

Added

• MCP: opt-in OAuth 2.0 resource-server profile for the remote HTTP profile. Setting PIPEFY_MCP_RS_RESOURCE_SERVER_URL activates it (no separate enable flag): the server validates an inbound Authorization: Bearer on every request (RS256 against the issuer's JWKS, issuer and expiry, optional audience), serves RFC 9728 protected-resource metadata, and returns 401 with a WWW-Authenticate challenge on a missing or invalid token. Token-validation knobs live in pipefy_auth.JwtValidationSettings (PIPEFY_JWT_ISSUER_URL, which defaults to the login issuer, plus PIPEFY_JWT_AUDIENCE / PIPEFY_JWT_VERIFY_AUDIENCE / PIPEFY_JWT_JWKS_URI); resource identity (PIPEFY_MCP_RS_RESOURCE_SERVER_URL, PIPEFY_MCP_RS_REQUIRED_SCOPES) stays in pipefy_mcp.ResourceServerSettings; PIPEFY_ALLOW_INSECURE_URLS covers both. The stdio profile is unchanged. Closes #301.
• MCP: the remote resource-server profile now acts on behalf of the validated caller (#302). An app-scoped runtime owns one shared client, and each request resolves its identity from its own inbound bearer (read from the request context), so one server serves concurrent callers as themselves rather than as a single startup identity. Credential resolution and its fail-fast move to the composition root (McpRuntime.for_profile / StartupIdentity.from_configured_credential); the stdio profile still runs as the one credential resolved at startup. Closes #302.
• SDK / MCP / CLI / auth: outbound Pipefy requests now carry client-identifying telemetry headers so the API can attribute traffic by surface and version. The GraphQL path (public, Interfaces, and Internal API endpoints) sends User-Agent: pipefy-sdk/<version> (<surface>), X-Client-Name: <surface>, and X-Client-Version: <version>, where <surface> is mcp, cli, or sdk. The surface is stamped programmatically at each entry point's composition root: the MCP server passes mcp, the CLI cli, and direct SDK use keeps the sdk default. It is a PipefyClient constructor argument rather than configuration, so it is never read from env or TOML and a caller cannot forge it. OAuth requests to the sign-in host carry User-Agent: pipefy-auth/<version>, X-Client-Name: auth, and X-Client-Version, identifying the auth component rather than an API surface. The header builder lives in pipefy_infra.telemetry so the SDK and auth packages share one User-Agent format. Closes #336.

Changed

• SDK: PyPI distribution renamed from pipefy-sdk to pipefy (the previous name is refused by PyPI as too similar to the unrelated pipefysdk). The import package is unchanged (pipefy_sdk); update installs and --with / uv add / pip install references accordingly. The distribution now builds as pipefy-<version>-*.whl.
• MCP (breaking): replaced the --remote launch flag with two explicit, orthogonal flags: --profile {local|remote} (env PIPEFY_MCP_PROFILE) and --transport {stdio|http} (env PIPEFY_MCP_TRANSPORT). The PIPEFY_MCP_REMOTE_MODE env var and the McpSettings.remote_mode field are removed; the default-deny remote-safe tool surface now follows profile == "remote". local (default) registers every tool; remote exposes only the remote-safe surface and validates an inbound bearer per request when a resource-server URL is set. Transport defaults from the profile (local->stdio, remote->http) and can be set explicitly to run local over loopback HTTP; remote over stdio is rejected (a per-request bearer has no stdio equivalent). The launch flags are resolved and validated once at startup by the new resolve_mcp_settings composition root, so argv outranks PIPEFY_MCP_*. The remote profile acts on behalf of the validated caller (see the request-scoped identity entry under Added); local runs as the one credential resolved at startup.
• SDK: services now receive GraphQL execution through their constructor instead of inheriting it. A narrow GraphQLExecutor protocol is the seam; HttpxGraphQLExecutor (formerly BasePipefyClient) is the only httpx/gql adapter, and PipefyClient builds one executor per endpoint via build_executors and injects them. Tests inject a fake executor rather than monkeypatching execute_query.
• SDK (breaking): the InternalApiClient class and its pipefy_sdk export are removed; the internal endpoint is now an HttpxGraphQLExecutor configured with internal_api_url and the [code=…][correlation_id=…] error formatter (build it via build_executors(settings, auth).internal). Service constructors no longer take auth, and RelationService/PortalService/MemberService/WebhookService/PipeConfigService now require their collaborators (no self-construct fallbacks).
• SDK: PipefyClient now constructs its internal API client during __init__ (built from the same settings and auth as every other endpoint client), so PipefyClient(settings, auth=...) is fully wired after construction. The set_internal_api_client methods on PipefyClient and PortalService are removed; pass the internal API client through the PortalService constructor when composing it directly. The PipefyClient.internal_api_available property is also removed: the internal API client is always present, so the property was always True.
• Auth (breaking): resolve_pipefy_auth now returns a ResolvedAuth sum type (StaticTokenAuth | ServiceAccountAuth | StoredSessionAuth) instead of an httpx.Auth. It parses the credential precedence chain into the tier that won, carrying that decision in the returned type: pattern-match or isinstance on the variant to branch on the tier and reach its fields. The new build_httpx_auth(resolved) constructs the transport httpx.Auth and is total over the sum. Callers that want the auth object in one step compose build_httpx_auth(resolve_pipefy_auth(...)).
• Plugin (Claude Code): renamed the OAuth slash command from /login to /pipefy-login (commands/login.md → commands/pipefy-login.md, frontmatter name: pipefy-login). The old name: login claimed the /login trigger and shadowed Claude Code's built-in /login (Claude account sign-in): selecting the built-in entry in the command picker still ran the Pipefy command, so Claude login was unreachable via /login while the plugin was installed. Closes #331.

Removed

• Auth (breaking): removed the unused CallableBearerAuth httpx adapter and its pipefy_auth export. It had no production construction site; RefreshableBearerAuth already covers the per-request-provider behaviour (re-calling the token provider each request) plus the 401 refresh-and-retry the stored-session path needs.
• Auth (breaking): removed pipefy_auth.tier_for and the internal _TIER_BY_AUTH_TYPE reverse-lookup table. They recovered a tier name from a built httpx.Auth by isinstance; the ResolvedAuth variants returned by resolve_pipefy_auth now carry the tier in their type, so branch on the variant (isinstance / match) instead.
• SDK / MCP / CLI (breaking): retired the PIPEFY_SERVICE_ACCOUNT_IDS env var and the service_account_ids setting (and its TOML key). This drops the advisory features it fed: the guard that blocked remove_member_from_pipe from removing a configured service account, the set_role write-permission warning, and the proactive cross-pipe membership check in validate_ai_agent_behaviors. The Pipefy API still authorizes every mutation, so removing a service account is recoverable via the Pipefy UI, and the cross-pipe membership requirement now surfaces as a bare PERMISSION_DENIED at call time (recover with get_pipe_members + invite_members). The field was an advisory, single-tenant shadow of a missing platform attribute and had no place in the remote multi-user profile.

v0.2.0-beta.4

[0.2.0-beta.4] - 2026-06-19

Added

• SDK / MCP / CLI: create_card accepts optional phase_id and passes title on GraphQL CreateCardInput (mutation createCard(input: $input)). MCP: interactive phase_id elicits start-form fields; non-empty fields are filtered against get_phase_fields(phase_id) and get_start_form_fields(pipe_id); agent seeding should set skip_elicitation=true. CLI: pipefy card create <pipe_id> --phase-id <id> (optional --title; emits title_warning when the API overrides the requested title).
• MCP / CLI: get_phase_allowed_move_targets lists valid destination phases for move_card_to_phase (GraphQL phase.cards_can_be_moved_to_phases). MCP returns normalized {phase_id, phase_name, allowed_phases}; CLI: pipefy phase targets <phase_id>.
• SDK / MCP / CLI: get_phase_cards_count reads native Phase.cards_count (MCP/CLI return {phase_id, phase_name, cards_count} via get_phase). get_phase_cards lists cards in a phase via Phase.cards pagination (first default 50, after, optional fields). CLI: pipefy phase count <phase_id>, pipefy phase cards <phase_id> --first --after.
• SDK / MCP / CLI: get_pipe selects cards_count on each workflow phase in phases[] (start-form intake stays on start_form_fields).
• SDK / MCP / CLI: report create/update/export paths preflight filter as nested ReportCardsFilter (operator + queries) in the SDK before GraphQL; reject naive top-level keys such as current_phase; coerce integer value scalars to strings. MCP and CLI surface the same error; CLI rejects invalid filters before auth.

Changed

• SDK / MCP / CLI (label color): create_label and update_label validate color as hex #RGB or #RRGGBB in the SDK before GraphQL (e.g. red is rejected with expected #RGB or #RRGGBB hex color, received 'red'). CLI rejects invalid color before auth.
• Docs / install: canonical GitHub repository is pipefy/ai-toolkit. Install snippets, install.sh, MCP setup links, Claude plugin metadata, and .mcp.json now point at github.com/pipefy/ai-toolkit (GitHub may still redirect older URLs).
• SDK: GraphQL HTTPXAsyncTransport now sets verify=True explicitly to pin TLS verification on, independent of httpx's default.
• Deps: upgraded pydantic to 2.13.4 (pydantic-core 2.46.4, pydantic-settings 2.14.1) across workspace packages.
• Deps: upgraded gql[httpx] to 4.0.0; gql() constants are GraphQLRequest objects (use .document for AST introspection in tests).

v0.2.0-beta.3

[0.2.0-beta.3] - 2026-06-03

Fixed

• SDK / MCP / CLI: AI automation create and update no longer require service-account credentials. generate_with_ai create/update were routed through the /internal_api endpoint and gated behind a service-account check, so a normal user session (stored session from pipefy auth login, or a --token bearer) could not create or update an AI automation, which broke install flows that provision an AI triage automation. Both now go through the public createAutomation / updateAutomation mutations under the caller's normal auth, matching the read, list, and delete paths. Closes #272.

Changed

• SDK: the AI automation surface moved from AiAutomationService to AutomationService.create_ai_automation / update_ai_automation, mirroring the existing create_send_task_automation pattern. AiAutomationService, ai_automation_queries.py, the ai_automation_available property, and set_ai_automation_service are removed; PipefyClient delegates to AutomationService. InternalApiClient stays (still used by delete_card_relation).

v0.2.0-beta.2.dev1

[0.2.0-beta.2.dev1] - 2026-06-02

Snapshot of dev HEAD ahead of v0.2.0-beta.2. See [Unreleased] above for the cumulative change list.

v0.2.0-beta.2

[0.2.0-beta.2] - 2026-06-02

Added

• MCP / CLI / SDK: get_automation_event_attributes lists official field_map.value tokens; create_automation preflights field_map destination fieldId values (numeric internal_id on action_repo_id) and card_moved + move_single_card transitions before GraphQL. Skill and docs/mcp/tools/automations-and-ai.md document create-only preflight and text-only move-transition errors (recovery via error message or get_phase_allowed_move_targets, not a valid_destinations envelope field).
• Installer: new POSIX install.sh at the repo root. One command installs the CLI + MCP server via uv tool install (resolving the latest GitHub Release tag via the GitHub API and discovering the wheel assets directly, no hardcoded package list), optionally adds skills via npx skills add, and writes the MCP server registration into the user's chosen client config (Cursor / Claude Desktop / Codex / Claude Code / none-mode print). Flags: --yes, --no-skills, --client <id>, --version <tag>, --prefix <dir>, --allow-root, --dry-run, --help. After install, pipefy-mcp-server is on PATH so client configs collapse to {"command": "pipefy-mcp-server"}. shellcheck-clean. Closes #231.
• SDK / Auth: shared filesystem-backed configuration via ~/.config/pipefy/config.toml (%APPDATA%\pipefy\config.toml on Windows; honours XDG_CONFIG_HOME and a PIPEFY_CONFIG_FILE=<path> override). Top-level TOML keys map to pydantic field names on AuthSettings and PipefySettings; precedence is init kwargs > env > .env > config.toml > defaults. The new pipefy-infra workspace package owns the loader (a schema-agnostic PydanticBaseSettingsSource subclass) and the path-discovery helpers; SDK and Auth depend on it symmetrically — neither imports the other. See docs/config.md for the schema.
• CLI: pipefy auth status [--json|-j] — reports auth source, identity, session expiry, and exit codes.
• CLI: pipefy auth logout — revokes the refresh token at the IdP and clears the stored session.
• Auth: AuthSettings.auth_url now defaults to the Pipefy production IdP when PIPEFY_AUTH_URL is unset — removes the previous PIPEFY_AUTH_URL is required exit-2 friction on pipefy auth login / pipefy auth logout and wires the MCP stored-session tier automatically after pipefy auth login. Override by setting PIPEFY_AUTH_URL=<url> to a non-prod IdP. Closes #233.
• Auth / CLI / MCP: opt-out for the keychain-backed stored-session tier. PIPEFY_DISABLE_STORED_SESSION=1 (or disable_stored_session=true in config.toml) makes AuthSettings.to_oidc_client() return None: tier resolution skips the keychain entirely (no backend discovery, no load_session probe), and pipefy auth login / pipefy auth logout refuse with exit code 2. PIPEFY_KEYCHAIN_BACKEND=file (or keychain_backend = "file") swaps the active keyring backend to keyrings.alt.file.PlaintextKeyring writing under ~/.config/pipefy/keyring.cfg (%APPDATA%\pipefy\keyring.cfg on Windows). Unblocks headless Linux without Secret Service and CI runners; the file stores credentials in plaintext on disk, so it's opt-in only. Closes #237.
• SDK / Auth: introduced PIPEFY_BASE_URL (default https://app.pipefy.com) that drives the four API endpoints (graphql_url, internal_api_url, interfaces_graphql_url, service_account_url) as pydantic @computed_field properties. Operators on non-prod environments set PIPEFY_BASE_URL=<host> once and all four endpoints follow. The OIDC issuer (PIPEFY_AUTH_URL, default https://signin.pipefy.com/realms/pipefy) remains a separate full-URL field because non-prod realm names don't follow a derivable convention. Closes #238.
• MCP: stored-session tier wired into ServicesContainer; setting PIPEFY_AUTH_URL after pipefy auth login now lets the MCP server reuse the keychain-backed session, with the refresh pre-warmed at startup so a stale or revoked session surfaces before the first tool call.
• Auth: reactive refresh-on-401 for the stored-session tier (RefreshableBearerAuth). When an API call returns 401, the SDK transport forces a refresh via ensure_fresh_session(force=True), persists any rotated tokens, and retries the request once before surfacing the error. Complements the eager refresh path for IdP-side revocation and mid-process token expiry. Closes #137.

Changed

• SDK / MCP / CLI: renamed the two service-account credential env vars for clarity and to remove the one-letter footgun against PIPEFY_AUTH_URL (interactive user-login issuer): PIPEFY_OAUTH_CLIENT → PIPEFY_SERVICE_ACCOUNT_CLIENT_ID, PIPEFY_OAUTH_SECRET → PIPEFY_SERVICE_ACCOUNT_CLIENT_SECRET. PIPEFY_OAUTH_URL is dropped without a renamed counterpart — the OAuth token endpoint now derives from PIPEFY_BASE_URL (see the Removed section). Closes #127.
• MCP / CLI: attachment uploads accept file_path only (local filesystem path, with ~ expansion). file_url and file_content_base64 are both gone — the MCP server runs as a local subprocess of the user's agent runtime, so any path the user can read is a valid source, and agents that hold generated bytes write them to a temp file and pass the path. file_name is optional (basename inferred from the path). The SDK exposes a single client.upload_attachment(attachment, organization_id=..., target=...) method backed by AttachmentService.upload_attachment, which owns the full choreography (file read → presigned URL → S3 PUT → field update) and enforces the 100 MiB cap as part of its internal LocalFile.read. Per-step wire helpers (create_presigned_url, upload_file_to_s3, extract_storage_path) are private to AttachmentService; the S3 PUT step is dispatched via an injectable S3Uploader protocol (HttpxS3Uploader is the default). Domain types: pipefy_sdk.Attachment (path-based, no upload methods), CardTarget / TableRecordTarget value objects, plus the AttachmentUploadError / AttachmentUploadResult / AttachmentUploadStep types previously in attachment_upload. See packages/mcp/AGENTS.md for the GATED:SELF_HOSTED convention that marks where URL ingestion would return under a future self-hosted profile.
• Infra: pipefy_infra.security widens the literal-IP / DNS-result blocklist to include multicast (224.0.0.0/4), reserved (240.0.0.0/4), and unspecified (0.0.0.0, ::) ranges in addition to the prior private / loopback / link-local set. Error messages updated to enumerate the blocked categories explicitly. The widening relies on stdlib ipaddress.IPv4Address.is_* / IPv6Address.is_* properties so IPv4-mapped IPv6 literals (::ffff:127.0.0.1) are also covered.
• Infra: security.validate_https_url and security.assert_url_is_host_root no longer take a derived_paths_hint parameter. Callers wrap the ValueError with their own caller-specific context if needed; the helpers stay caller-agnostic.
• Auth: PIPEFY_KEYCHAIN_BACKEND env / TOML values are now normalized with .strip().lower() so AUTO, File, etc. match the Literal["auto", "file"] constraint instead of failing with a cryptic Literal-validation error. Previous behavior accepted only the exact lower-case spelling.
• Auth: PIPEFY_AUTH_URL now rejects query strings and fragments (path is allowed for Keycloak-style realm URLs). Previously, only validate_https_url's scheme + literal-IP check ran on auth_url, so a stray ? or # from operator copy-paste would pass settings construction and corrupt the downstream .well-known/openid-configuration concatenation.
• Infra: security.assert_url_is_host_root rejects repeated-slash paths (//, ///) that the previous inline path.strip('/') check accepted. Deploy configs with PIPEFY_BASE_URL=https://app.pipefy.com// (extra slash) now fail at startup; trim to a single trailing slash or none.
• Infra: security.assert_hostname_resolves_to_public_ips raises ValueError instead of silently passing when socket.getaddrinfo returns an empty address list. UnicodeError from IDN encoding (e.g. labels > 63 chars) is now caught alongside socket.gaierror and surfaced as a ValueError with the same Could not resolve hostname shape.
• Auth / SDK: every PIPEFY_* env var is validated against a semantically meaningful pattern at settings construction. URL env vars (PIPEFY_BASE_URL, PIPEFY_AUTH_URL) require https?:// plus non-whitespace; credential fields (PIPEFY_TOKEN, PIPEFY_SERVICE_ACCOUNT_CLIENT_ID, PIPEFY_SERVICE_ACCOUNT_CLIENT_SECRET, PIPEFY_AUTH_CLIENT_ID) reject leading / trailing whitespace; PIPEFY_ORG_ID must be an ASCII numeric string. There is no longer an empty-string opt-out for any tier — unset the variable to fall back to the default, or use PIPEFY_DISABLE_STORED_SESSION=1 to turn the stored-session tier off explicitly (see the Added entry above).
• SDK / MCP / CLI: renamed --graphql-url CLI flag to --base-url to match the new env-var shape.
• Install snippets: every git+https://github.com/gbrlcustodio/pipefy-mcp-server URL in the root README.md, packages/mcp/README.md, packages/cli/README.md, the shipping .mcp.json, and commands/install.md now pins @latest — a moving git tag the release flow updates to point at the most recent release. Users get release-stable installs by default; @v0.2.0-beta.1-style pins remain available for reproducibility. The release process documented in RELEASE.md now includes git tag -f latest <version> && git push --force-with-lease origin latest as the step that rolls new installs forward.

Deprecated

• SDK: legacy PIPEFY_OAUTH_* env vars still resolve to the new service_account_* fields via an alias shim, with a one-shot stderr deprecation warning per legacy key. The aliases will be removed in a later 0.2.0-beta.x release (carrying an explicit breaking-change callout). See [`...

v0.2.0-beta.1

[0.2.0-beta.1] - 2026-05-18

Monorepo Pipefy Labs public beta on the v0.2.0-beta.* line (GitHub Release + wheels only; no PyPI until v1.*). Tag v0.2.0-beta.1 matches __version__ in all workspace packages per RELEASE.md.

Added

• CLI: added MCP-parity commands for core workflow domains: pipe, phase,
  field, table, record, label, webhook, relation, and member.
• CLI: added post-v0.1 parity domains: attachment, field-condition,
  email, audit, automation, introspect, graphql, agent,
  ai-automation, usage, report-pipe, report-org, export, and org.
• MCP / CLI: shared SDK facade covers attachment upload, automation
  preflight, field-condition normalization, AI prompt and behavior validation,
  report export streaming, and service-account guard helpers.
• CLI: pipefy skills list and pipefy skills show <name> for browsing the bundled
  starter pack (8 high-impact Pipefy workflows), with YAML frontmatter parsing for
  descriptions.
• Skills: skills/ catalog with authoring guide, contributing rules, and
  skills-lint.yml CI (frontmatter, starter-pack bundle drift, MCP/CLI reference lint,
  and pipefy skills list smoke).
• Docs: docs/MIGRATION.md cutover guide for existing pipefy-mcp-server users.
• Tooling: scripts/sync_starter_pack.py copies canonical starter-pack SKILL.md
  files into packages/cli/src/pipefy_cli/skills/; use --check in CI or before release.
• CLI: introduce pipefy-cli workspace package with pipefy entry point.
• CLI: pipefy card get <id> (mirrors MCP get_card) with --json / Rich rendering.
• CLI: OAuth client-credentials auth (PIPEFY_OAUTH_*) and --token / PIPEFY_TOKEN static bearer override; auth precedence flag > env > ~/.config/pipefy/config.toml.
• CLI: --graphql-url and --allow-insecure-urls global flags; same SSRF policy as MCP.
• CLI: shell completion via pipefy --install-completion bash|zsh.
• SDK: optional bearer_token= constructor on PipefyClient and StaticBearerAuth in base_client (transport auth path used by the CLI --token / PIPEFY_TOKEN).

Changed

• Docs: MCP tool reference moved to docs/mcp/tools/; added docs/README.md, docs/mcp/README.md, docs/cli/README.md, and docs/sdk/README.md as surface-oriented entry points. docs/setup.md and docs/parity.md paths unchanged for stable links.
• SDK: PyPI distribution renamed from pipefy-ai-sdk to pipefy-sdk (import package remains pipefy_sdk). Update installs and uv add / pip install references accordingly.
• CLI / MCP: Creating a traditional automation with card_moved + move_single_card runs SDK move-transition preflight first, returning a clear validation error when the destination phase is unreachable from the source phase (instead of opaque GraphQL failures).
• Internal: repository reorganized as a uv workspace; pipefy-mcp-server distribution and runtime behavior unchanged.

Fixed

• CLI: pipefy agent update resolves slug-style fieldId values in behaviors for error-path enrichment the same way as the happy path (via PipefyClient.update_ai_agent), so RECORD_NOT_SAVED diagnostics do not falsely blame slug tokens as unknown pipe fields.
• CLI / MCP: field-condition create / update accept legacy actionId: "hidden" on condition actions; the SDK normalizes to hide before mutations.
• SDK: PipeConfigService.update_phase_field accepts optional phase_id / pipe_id and resolves a slug-like field_id to the field's uuid (injected as input.uuid while the slug stays as input.id, matching Pipefy's UpdatePhaseFieldInput contract). The pipe-wide lookup runs phase fetches concurrently via asyncio.gather; partial phase-fetch failures raise an actionable ValueError instead of returning an ambiguous match. Surfaced through MCP update_phase_field(phase_id=…, pipe_id=…) and CLI pipefy field update --extra '{"phase_id":"…"}'.
• MCP: delete_phase_field preview now enumerates dependents.field_conditions even when the rule only references the field in expression field_address (not just actions[].phaseFieldId); the condition tree walker has a defensive depth cap of 16.
• SDK / MCP: PipefyClient.get_automation_logs_by_repo short-circuits to an empty page when the pipe has no automations (was returning MULTIPLE_INVALID_INPUT: Automation_ids can't be blank from the API).
• SDK / MCP: invite_members validates each row with a new MemberInvite Pydantic model (EmailStr + non-blank role_name, lowercase normalization, extra="forbid") and raises a single-line ValueError pointing at the offending field. MCP surfaces it as INVALID_ARGUMENTS.
• SDK: ai_preflight.validate_ai_automation_prompt_sdk flags overlap when the same %{internal_id} appears both in the prompt and in field_ids, in English, citing the API rejection message.
• MCP: find_records returns the unified envelope pagination={has_more, end_cursor, page_size} (snake_case) when the unified envelope flag is on, matching get_table_records.
• MCP / CLI docs: create_card, create_table_record, clone_pipe, and create_field_condition docstrings clarify title-derivation quirks, async clone phases, and the phaseFieldId discovery path.
• SDK: MemberInvite lives at pipefy_sdk.MemberInvite (re-exported in the top-level __all__); slug_like_field_token / looks_like_uuid_token extracted to pipefy_sdk.utils.field_tokens for reuse across services.

Removed

Public beta: 128 MCP tools for Pipefy

Public beta of the Pipefy MCP server — an open-source Model Context Protocol server that lets AI agents drive end-to-end Pipefy workflows (pipes, cards, tables, relations, reports, automations, AI agents, observability, and more).

🧪 Beta · pre-release, built in public. Community project for developer workflows — not Pipefy's official or supported integration for external enterprise use. The tool surface is feature-complete for this milestone, but names/arguments/response shapes may still change before the stable v0.1.0.
Feedback & issues: https://github.com/gbrlcustodio/pipefy-mcp-server/issues · dev@pipefy.com

Highlights

• 128 MCP tools across nine surface areas (canonical names in PIPEFY_TOOL_NAMES, src/pipefy_mcp/tools/registry.py):
• Pipes & cards (37) · Database tables (17) · Relations (8) · Reports (17)
• Automations & AI (22) · Observability (10) · Members, email & webhooks (11)
• Organization (1) · Introspection (5)
• Agent-friendly responses: unified success envelope and shared pagination bounds on list/search tools; structured error envelopes (INVALID_ARGUMENTS, enriched GraphQL errors) across domains.
• Safer writes: destructive-action previews with dependents (labels, phases, phase fields), two-step deletes, and MCP elicitation for confirmation flows.
• Schema introspection tools with depth control and raw GraphQL execution for advanced agent reasoning.
• Attachment uploads for cards and table records via presigned S3.
• Field slug → numeric ID resolution for AI agent behaviors, so agents can author behavior metadata using human-readable slugs.
• Authentication via Pipefy Service Account OAuth2 (httpx-auth), shared across GraphQL and direct HTTP clients.
• Parallel-safe transport: per-request GraphQL transport so concurrent tool calls don't share a single connection.
• Configurable client timeout on the Pipefy HTTP client.
• find_cards tool plus include_fields on get_card / get_cards for field-rich reads when agents need them.
• update_comment / delete_comment card-comment tools.

Getting started

• Python 3.11+ and a Pipefy Service Account Token (Admin Panel → Service Accounts)
• uv sync · cp .env.example .env · or run ./bootstrap.sh
• MCP client setup (Cursor / Claude Desktop / Claude Code): see docs/setup.md
• Tool reference: per-category docs under docs/tools/ and cross-cutting rules in docs/tools/cross-cutting.md

Known limitations

• Beta surface — tool names, arguments, and response shapes may still receive breaking changes before the stable v0.1.0 cut. Pin the tag if you need stability while we collect feedback.
• Complex Pipefy field types may need agent-side mapping help; please open an issue with the field type and pipe context if you hit one.

Contributing

We're building this in public and beta feedback is exactly what we need — new-tool requests, field-mapping reports, and bug reports very welcome. See the "Contributing" section in the README.

Full changelog: https://github.com/gbrlcustodio/pipefy-mcp-server/commits/v0.1.0-beta