API reference

This page lists the import path for every public name in loomflow, grouped by purpose, so you can find where something lives. Each group links to the page that explains how to use it. For full parameter-by-parameter references see Agent, Workflow and the architecture classes.

The package ships a py.typed marker, so IDEs (Pylance, Pyright, mypy) get autocomplete and go-to-definition on every public symbol. If your IDE doesn’t pick this up, see Common pitfall: stale install.

Agent


from loomflow import Agent, Tuning
from loomflow import ApprovalDecision, GoalStopHook, StopHook, StopHookResult

The model-driven loop. See What is an Agent and the Agent reference.

Tuning groups the rarely-touched knobs (retry_policy, secrets, auto_consolidate, response_tone, stop_hooks, max_stop_hook_iterations, checkpoint, tool search, memory budget, and the tool-transcript / auto-compact settings) into one optional dataclass: pass Agent(..., tuning=Tuning(...)). See Tuning fields.

ApprovalDecision is the rich return type for approval handlers. GoalStopHook is what Agent(run_until=...) installs; StopHook / StopHookResult let you write your own.

Workflow


from loomflow import END, START, Workflow, WorkflowResult, step

Graphs of steps you define in code. See Workflow and the Workflow reference.

Architectures


from loomflow import Architecture, ReAct
from loomflow.architecture import ActorCritic, Blackboard, BlackboardArchitecture, BlackboardEntry, MultiAgentDebate, Plan, PlanAndExecute, PlanStep, Reflexion, ReWOO, ReWOOPlan, ReWOOStep, ReWOOStepResult, Router, RouterRoute, SelfRefine, StepResult, Supervisor, Swarm, Handoff, ThoughtNode, TreeOfThoughts, AgentSession, Dependencies, resolve_architecture
from loomflow.team import Team, run_architecture

See Architectures.

Models


from loomflow import EchoModel, ScriptedModel, ScriptedTurn, Image, count_tokens
from loomflow.model import AnthropicModel, FallbackModel, LiteLLMModel, OpenAIModel, default_fall_on
from loomflow.model.retrying import RetryingModel

See Models · Fallback chains + timeouts · Vision.

Context management


from loomflow import context_window_for, maybe_auto_compact, snip_messages

The helpers behind Agent(snip_window=) and Agent(auto_compact_at_tokens=), exposed for custom architectures. See snip_window and auto_compact_at_tokens.

Memory


from loomflow import HashEmbedder, InMemoryMemory, resolve_memory
from loomflow.memory import ChromaFactStore, ConsolidationWorker, Consolidator, FactStore, InMemoryFactStore, PostgresFactStore, RedisFactStore, SqliteFactStore
from loomflow.memory.auto_extract import AutoExtractMemory
from loomflow.memory.chroma import ChromaMemory
from loomflow.memory.embedder import CohereEmbedder, OpenAIEmbedder, VoyageEmbedder
from loomflow.memory.lazy import LazyMemory
from loomflow.memory.postgres import PostgresMemory
from loomflow.memory.redis import RedisMemory
from loomflow.memory.sqlite import SqliteMemory
from loomflow.memory.vector import VectorMemory

See Memory.

Vector stores and loaders


from loomflow.vectorstore import ChromaVectorStore, FAISSVectorStore, InMemoryVectorStore, PostgresVectorStore, SearchResult, VectorStore
from loomflow.loader import Chunk, Document, MarkdownChunker, RecursiveChunker, SentenceChunker, TokenChunker, chunk, load

See RAG · Loaders · Chunkers.

Tools


from loomflow import LivingPlan, LivingPlanStep, Tool, tool, web_tool
from loomflow.tools import InProcessToolHost, PathEscapeError, bash_tool, default_workdir, edit_tool, filesystem_tools, find_tool, grep_tool, ls_tool, read_tool, write_tool
from loomflow.tools import CodeExecutor, SubprocessExecutor, make_code_mode_tools
from loomflow.tools.plan import get_active_plan, make_plan_tools, make_recall_past_plans_tool

See Tools · Built-in tools · Code mode · Living plan.

Workspace


from loomflow import (
    InMemoryWorkspace,
    LocalDiskWorkspace,
    Workspace,
    WorkspaceMembership,
    resolve_workspace,
)
from loomflow.workspace import (
    Note,
    NoteKind,
    NoteMatch,
    NoteSummary,
    NoteVersion,
    PruneResult,
    make_workspace_tools,
    workspace_prompt_section,
)

See Workspace · Workspace lifecycle.

Skills


from loomflow.skills import Skill, SkillError, SkillMetadata, SkillRegistry, SkillSource

See Skills.

MCP


from loomflow.mcp import MCPClient, MCPRegistry, MCPServerSpec

See MCP.

Runtime


from loomflow import InProcRuntime
from loomflow.runtime import Checkpoint, CheckpointMeta, InMemoryJournalStore, JournaledRuntime, JournalStore, PostgresJournalStore, PostgresRuntime, SqliteJournalStore, SqliteRuntime, journal_scope, resolve_runtime

See Runtime · Checkpoints.

Security


from loomflow import AllowAll, AuditLog, FullTranscriptAuditLog, HookRegistry, InMemoryAuditLog, Mode, NoSandbox, OSSandbox, StandardPermissions
from loomflow.security import AuditProblem, AuditVerification, DictSecrets, EnvSecrets, FileAuditLog, FilesystemSandbox, PerUserPermissions, SubprocessSandbox, verify_chain, verify_signature

See Security.

Guardrails


from loomflow.guardrails import GuardedStream, Guardrail, GuardrailOutcome, GuardrailTrigger, GuardVerdict, InjectionGuard, ModerationGuard, PIIGuard, RegexGuard, apply_guardrails

Not re-exported from the top-level package. See Guardrails.

Governance


from loomflow import BudgetConfig, NoBudget, StandardBudget, TokenBucketRateLimiter
from loomflow.governance import RateLimiter, RetryPolicy, classify_model_error, compute_backoff

See RetryPolicy + error taxonomy · Per-user budget caps · Rate limiting.

Decisions


from loomflow.decisions import Box, Choice, DecisionApprovalPolicy, DecisionGuard, DecisionModel, Decisions, JevModel, LLMDecisionModel, Noul, Question, Score, ScriptedDecisions

Typed answers with probabilities for the framework’s internal judgments (which route, is the goal met, is this call risky) without a full LLM call. See Decisions.

Observability


from loomflow import NoTelemetry
from loomflow.observability import (
    InMemoryTelemetry,
    ConsoleTelemetry,
    FileTelemetry,
    MultiTelemetry,
    OTelTelemetry,
    CapturedSpan,
    CapturedMetric,
)
from loomflow.graph import AgentGraph, build_graph, write_graph

See OpenTelemetry · AgentGraph.

Studio


import loomflow.studio
from loomflow.studio import dataset, disable, enable, flush, is_enabled, prune, score, session
from loomflow.studio.web import create_app

enable(path, project=, sample_rate=, capture_content=, redact=, keep_days=) records every run; session(project=, tags=, metadata=, enabled=) scopes it; score(result, key, value) attaches a score; prune(days) deletes old runs; dataset(name) loads a dataset built in Studio as an eval Dataset; create_app(path, read_only=, allowed_hosts=, agents=) is the loom studio web app (agents= are the agents its page may run evals on). Environment: LOOMFLOW_STUDIO, LOOMFLOW_STUDIO_PROJECT, LOOMFLOW_STUDIO_KEEP_DAYS. See Studio.

Evals


from loomflow.eval import Case, CaseResult, Dataset, EvalHarness, EvalReport, LLMJudge

See Evals.

Serving and A2A


from loomflow.serve import create_app
from loomflow.a2a import A2AClient, AgentCard, serve_a2a

See Serving over HTTP · A2A protocol.

Data lineage


from loomflow.core.types import CertifiedValue
from loomflow.data import FreshnessPolicy, LineagePolicy, check_freshness, check_lineage, require_freshness, require_lineage

CertifiedValue wraps a value with its source, fetch time, expiry and lineage. The policies check it and raise FreshnessError / LineageError (see Errors).

Core types


from loomflow import BudgetStatus, Episode, EpisodeMatch, Event, EventKind, Fact, MemoryBlock, MemoryExport, MemoryProfile, Message, PermissionDecision, Role, RunResult, ToolCall, ToolDef, ToolResult, Usage
from loomflow.core.types import AuditEntry, ModelChunk, PromptCacheConfig, Span, ToolEvent

PromptCacheConfig is what the Agent(prompt_caching=) kwarg resolves to. Usage carries cached_input_tokens and RunResult carries cached_tokens_in, plus cache_write_tokens on both. See Prompt caching.

Protocols


from loomflow import Budget, Embedder, HookHost, Memory, Model, Permissions, Runtime, Sandbox, Secrets, Telemetry, ToolHost

Implement any of these to add a new backend without forking the framework.

RunContext


from loomflow import IsolationWarning, RunContext, get_run_context, set_run_context

See Run vs Session vs RunContext.

Errors


from loomflow import BudgetExceeded, ConfigError, LoomError, OutputValidationError, RateLimitExceeded, RunTimeout
from loomflow.core.errors import AuthenticationError, CancelledByUser, ContentFilterError, FreshnessError, InvalidRequestError, LineageError, MCPError, MemoryStoreError, ModelError, PermanentModelError, PermissionDenied, RateLimitError, RuntimeJournalError, SandboxError, ToolError, TransientModelError
from loomflow.model._effort import EffortNotSupportedError
from loomflow import LoomDeprecationWarning

See Errors.

Identity helpers


from loomflow import new_id
from loomflow.core.ids import deterministic_hash

new_id() returns a fresh ULID. deterministic_hash(obj) returns a stable hash of any JSON-serializable object. The framework uses it for journal keys; it is also useful in custom architectures.

Version


from loomflow import __version__

__version__ is the installed version string. These docs describe loomflow 0.17.