ChainWeaver

Observe the tool paths your agent repeats. Compile them into typed, deterministic flows. Replace the LLM-in-the-loop with governed, auditable execution.

The moat — observe → compile → replace. Point ChainWeaver at the tool paths your agent already repeats. ChainAnalyzer maps every schema-compatible chain among your tools; you compile the ones worth keeping into typed Flow objects; and FlowExecutor replaces the per-step LLM round-trips with deterministic, schema-validated execution — no model in the loop. You compile the path the analyzer surfaces instead of hand-wiring it.

Governance for tool flows. Typed I/O at every step, file-serializable flows, schema-drift detection, determinism attestation, property fuzzing, and structured audit traces — disciplined, auditable, portable deterministic execution.

Quantified and reproducible. In the repo's benchmark report, compiled flows show 0% data corruption versus 61–96% for naive LLM-in-the-loop chaining, and avoid ~$0.06 of LLM spend per 10-step flow. Regenerate it yourself with python benchmarks/report.py. Saving LLM calls is a consequence — not the headline.

from chainweaver import Tool, Flow, FlowStep, FlowRegistry, FlowExecutor
# (NumberInput, ValueOutput, double_fn defined in full example below)

# 1. Wrap any function as a schema-validated Tool
double = Tool(name="double", description="Doubles a number.",
              input_schema=NumberInput, output_schema=ValueOutput, fn=double_fn)
# 2. Wire tools into a Flow
flow = Flow(name="calc", description="Double a number.",
            steps=[FlowStep(tool_name="double", input_mapping={"number": "number"})])
# 3. Register and execute — zero LLM calls
registry = FlowRegistry()
registry.register_flow(flow)
executor = FlowExecutor(registry=registry)
executor.register_tool(double)
result = executor.execute_flow("calc", {"number": 5})
# result.final_output → {"number": 5, "value": 10}

See the full example below or run python examples/simple_linear_flow.py

Installation · Why ChainWeaver? · Is this for me? · Quick Start · Playground · Architecture · Docs site · Roadmap

See it in 30 seconds

The problem. Your agent keeps doing the same path — search → extract → validate → format — but on every single turn it round-trips through the LLM between each tool call to "decide" what to do next. That's four model calls to execute one deterministic operation.

Before — naive agent loop, 4 model-mediated decisions:

turn 1   ─►  LLM("plan")    ─►  search(query)         ─► 12 results
turn 2   ─►  LLM("next?")   ─►  extract(results)      ─► 8 facts
turn 3   ─►  LLM("next?")   ─►  validate(facts)       ─► 7 facts
turn 4   ─►  LLM("next?")   ─►  format(facts)         ─► answer
                                                          ⏱  ~6 s, 4 LLM calls

After — same path, compiled once into a named ChainWeaver flow:

turn 1   ─►  LLM("plan")    ─►  search_summarize_flow(query)
                                  └─ search ─► extract ─► validate ─► format
                                                          ⏱  ~1 s, 1 LLM call

The agent still decides which flow to invoke (that part stays open-ended). The four tool calls inside the flow no longer round-trip through the model — FlowExecutor runs them with strict Pydantic validation between every step and zero LLM involvement.

Copy-paste quick path:

pip install 'chainweaver[yaml]'
python examples/simple_linear_flow.py

The summary below is a condensed view of the real ExecutionResult the script produces — the actual stdout also includes per-step timestamps and the executor's structured step log, but the values, the step order, and the final output are exactly what you get on disk:

flow=double_add_format success=True
final_output={'number': 5, 'value': 20, 'result': 'Final value: 20'}
step 0 double          {'value': 10}
step 1 add_ten         {'value': 20}
step 2 format_result   {'result': 'Final value: 20'}

Three tool calls, no LLM in the loop, fully reproducible from examples/double_add_format.flow.yaml. Jump to the Quick Start for the Python version, or to the Command-line interface for the no-Python path.

Why ChainWeaver?

When an LLM-powered agent routes tools together — fetch_data → transform → store — a common pattern is to insert an LLM call between every step so the model can "decide" what to do next.

User request
    │
    ▼
LLM call ──► Tool A
    │
    ▼
LLM call ──► Tool B
    │
    ▼
LLM call ──► Tool C
    │
    ▼
Response

For flows that are fully deterministic (the next step is always the same given the previous output) these intermediate LLM calls add:

• Latency — each round-trip costs hundreds of milliseconds.
• Cost — every call consumes tokens and credits.
• Unpredictability — a language model might route differently on each invocation.

ChainWeaver compiles deterministic multi-tool flows into executable flows that run without any LLM involvement between steps:

User request
    │
    ▼
FlowExecutor ──► Tool A ──► Tool B ──► Tool C
    │
    ▼
Response

Think of it as the difference between an interpreter and a compiler:

How is this different from LangChain / LangGraph / Prefect / Dagster / Temporal?

Short answer: those frameworks each make a different design choice that's right for their own audience. ChainWeaver makes one specific trade-off — no LLM calls between steps, enforced at the framework level — and aligns the rest of the design (Pydantic-validated I/O, file-serializable flows, no server) around it.

See docs/comparisons.md (also on the hosted site) for the full matrix — version pins, citations to each alternative's own docs, and a "when to pick which" guide. Re-evaluated on each minor release of any of the projects above.

Is this for me?

ChainWeaver is built for one specific shape of problem. The full fit/non-fit page covers the nuances; the short version:

Use ChainWeaver when

• The flow is predictable — you can name the next tool from the previous output without asking a model to decide.
• Determinism matters — same input must produce the same output, same execution path, same trace.
• You want strict schemas, audit-grade traces, and zero LLM calls between deterministic steps.

Don't use ChainWeaver when

• Every step requires open-ended reasoning to pick the next one (use an agent framework: LangGraph, the OpenAI / Anthropic SDK tool-use loops).
• You need a general workflow engine for scheduled / durable jobs across time (use Prefect, Dagster, or Temporal).
• You expect the executor to call an LLM. It deliberately doesn't.

The framework comparison table above (and the full one-paragraph-per-tool matrix in docs/comparisons.md) covers how ChainWeaver relates to its neighbours — LangChain, LangGraph, Prefect, Dagster, and Temporal.

For the correctness argument behind the design, see docs/data-integrity.md.

Part of the Weaver Stack

ChainWeaver is the deterministic multi-step tool execution layer of the Weaver Stack — a family of small, composable SDKs that share weaver-spec's SelectableItem routing contract. On the request path a router picks which capability to invoke, ChainWeaver runs the deterministic tool path behind it, and downstream layers gate and guard the call:

flowchart LR
    req([Request]) --> ctx[contextweaver<br/>context assembly]
    ctx --> cw[<b>ChainWeaver</b><br/>deterministic flow execution]
    cw --> ak[agent-kernel<br/>capability gating]
    ak --> af[agentfence<br/>runtime guardrails]
    subgraph adjacent [Adjacent · use any subset]
        vg[vibeguard]
        lw[lessonweaver]
        se[skdr-eval]
    end

Use standalone or together. Each layer stands on its own — ChainWeaver's base install has no hard dependency on any sibling and works fully standalone. Real interop runs through the chainweaver[weaver-stack] extra, which pins the published weaver-contracts package: ChainWeaver consumes its SelectableItem / RoutingDecision / CapabilityToken types directly, so a router can hand a routing decision straight to resolve_flow_from_routing_decision() for deterministic execution. See the runnable Weaver Stack golden path (issue #234).

ChainWeaver does not replace an agent framework. It is meant to be called from one — see the LangGraph recipe (issue #205) and the OpenAI Agents SDK recipe (issue #206) for the canonical integration patterns.

For host-level expectations (when to invoke, how to store traces, side-effect tools, MCP parity), see the Runtime responsibilities page.

Installation

pip install chainweaver                  # base install — no extras
pip install 'chainweaver[yaml]'          # most common — needed for .flow.yaml files
pip install 'chainweaver[yaml,otel,mcp]' # combine extras with commas

The base install pulls only five runtime dependencies (deepdiff, packaging, pydantic, tenacity, typer) and has no transitive LLM SDK pinned. Pick extras for the integrations you actually use:

Package metadata (pyproject.toml) publishes URLs for the documentation, the source, the changelog, and the issue tracker, so pip show chainweaver and the PyPI sidebar point users to the right place.

Quick Start

Define tools, build a flow, and execute it

from pydantic import BaseModel
from chainweaver import Tool, Flow, FlowStep, FlowRegistry, FlowExecutor

# --- 1. Declare schemas ---

class NumberInput(BaseModel):
    number: int

class ValueOutput(BaseModel):
    value: int

class ValueInput(BaseModel):
    value: int

class FormattedOutput(BaseModel):
    result: str

# --- 2. Implement tool functions ---

def double_fn(inp: NumberInput) -> dict:
    return {"value": inp.number * 2}

def add_ten_fn(inp: ValueInput) -> dict:
    return {"value": inp.value + 10}

def format_result_fn(inp: ValueInput) -> dict:
    return {"result": f"Final value: {inp.value}"}

# --- 3. Wrap as Tool objects ---

double_tool = Tool(
    name="double",
    description="Takes a number and returns its double.",
    input_schema=NumberInput,
    output_schema=ValueOutput,
    fn=double_fn,
)

add_ten_tool = Tool(
    name="add_ten",
    description="Takes a value and returns value + 10.",
    input_schema=ValueInput,
    output_schema=ValueOutput,
    fn=add_ten_fn,
)

format_tool = Tool(
    name="format_result",
    description="Formats a numeric value into a human-readable string.",
    input_schema=ValueInput,
    output_schema=FormattedOutput,
    fn=format_result_fn,
)

# --- 4. Define the flow ---

flow = Flow(
    name="double_add_format",
    description="Doubles a number, adds 10, and formats the result.",
    steps=[
        FlowStep(tool_name="double",        input_mapping={"number": "number"}),
        FlowStep(tool_name="add_ten",       input_mapping={"value": "value"}),
        FlowStep(tool_name="format_result", input_mapping={"value": "value"}),
    ],
)

# --- 5. Execute ---

registry = FlowRegistry()
registry.register_flow(flow)

executor = FlowExecutor(registry=registry)
executor.register_tool(double_tool)
executor.register_tool(add_ten_tool)
executor.register_tool(format_tool)

result = executor.execute_flow("double_add_format", {"number": 5})

print(result.success)       # True
print(result.final_output)  # {'number': 5, 'value': 20, 'result': 'Final value: 20'}

for record in result.execution_log:
    print(record.step_index, record.tool_name, record.outputs)
# 0 double {'value': 10}
# 1 add_ten {'value': 20}
# 2 format_result {'result': 'Final value: 20'}

You can also run the bundled examples directly:

python examples/simple_linear_flow.py   # simple arithmetic flow
python examples/etl_flow.py             # ETL flow: fetch → validate → normalize → enrich → store
python examples/mcp_search_flow.py      # MCP-style search → extract → format flow
python examples/naive_vs_compiled.py    # timing comparison: naive LLM calls vs ChainWeaver flow
python examples/coding_agent_pr_review.py    # deterministic PR-review checklist
python examples/coding_agent_changelog.py    # changelog generation workflow template
python examples/coding_agent_debug_log.py    # debug-log triage workflow template
python examples/mcp_style_before_after_demo.py        # before/after MCP-style flow demo
python examples/release_readiness_flow/release_readiness.py  # deterministic release-readiness gate
python examples/skdr_policy_eval_flow.py              # offline policy-evaluation workflow template
python examples/integrations/langgraph_node.py        # call a flow from a LangGraph node (needs chainweaver[langgraph])
python examples/integrations/openai_agents_tool.py    # expose a flow as an OpenAI Agents SDK tool (needs chainweaver[openai-agents])

The hosted docs also include a cookbook with paired scripts under examples/cookbook/, plus framework recipes and workflow templates (LangGraph, OpenAI Agents SDK, release-readiness, policy evaluation).

With the @tool decorator

The @tool decorator eliminates boilerplate by introspecting type hints to auto-generate input schemas:

from pydantic import BaseModel
from chainweaver import tool, Flow, FlowStep, FlowRegistry, FlowExecutor

class ValueOutput(BaseModel):
    value: int

class FormattedOutput(BaseModel):
    result: str

@tool(description="Doubles a number.")
def double(number: int) -> ValueOutput:
    return {"value": number * 2}

@tool(description="Adds ten.")
def add_ten(value: int) -> ValueOutput:
    return {"value": value + 10}

@tool(description="Formats the result.")
def format_result(value: int) -> FormattedOutput:
    return {"result": f"Final value: {value}"}

flow = Flow(
    name="double_add_format",
    description="Doubles a number, adds 10, and formats the result.",
    steps=[
        FlowStep(tool_name="double",        input_mapping={"number": "number"}),
        FlowStep(tool_name="add_ten",       input_mapping={"value": "value"}),
        FlowStep(tool_name="format_result", input_mapping={"value": "value"}),
    ],
)

registry = FlowRegistry()
registry.register_flow(flow)

executor = FlowExecutor(registry=registry)
executor.register_tool(double)
executor.register_tool(add_ten)
executor.register_tool(format_result)

result = executor.execute_flow("double_add_format", {"number": 5})
print(result.final_output)  # {'number': 5, 'value': 20, 'result': 'Final value: 20'}

Decorated tools are also directly callable:

print(double(number=5))  # {'value': 10}

See examples/decorator_tool.py for a runnable before/after comparison.

With FlowBuilder

FlowBuilder provides a fluent, chainable API as a more Pythonic alternative to constructing Flow objects directly. It produces an identical Flow — it is syntax sugar, not a replacement:

from chainweaver import FlowBuilder

flow = (
    FlowBuilder("double_add_format", "Doubles a number, adds 10, and formats.")
    .step("double", number="number")
    .step("add_ten", value="value")
    .step("format_result", value="value")
    .build()
)

• .step(tool_name, **mapping) — adds a step; string values are context-key lookups, non-string values are literal constants, no kwargs = full-context passthrough.
• .step_from(flow_step) — appends a pre-built FlowStep for interop.
• .with_input_schema(Model) / .with_output_schema(Model) — optional flow-level Pydantic schema declarations.
• .with_trigger(conditions) — optional free-form trigger metadata.
• .build() — returns a validated Flow; raises FlowBuilderError if name or description is missing.

Interactive playground

Want to try ChainWeaver without installing anything locally? The playground/ directory ships a Streamlit app that lets you pick a pre-loaded flow, edit its JSON input, run it, and watch the step-by-step, LLM-free execution trace with a Mermaid diagram — the same FlowExecutor the library ships.

pip install -r playground/requirements.txt
streamlit run playground/app.py

It ships three example flows (arithmetic, a data flow, and an MCP-style search), produces shareable ?share=<token> links that round-trip a run through the URL, and is fully stateless so it deploys to Streamlit Community Cloud with no backend. See playground/README.md for local-run and deployment instructions.

Architecture

chainweaver/
├── __init__.py       # Public API
├── builder.py        # FlowBuilder — fluent API for flow construction
├── compat.py         # schema_fingerprint, check_flow_compatibility
├── compiler.py       # compile_flow — static schema flow validation
├── decorators.py     # @tool decorator for zero-boilerplate tool definition
├── tools.py          # Tool — named callable with Pydantic schemas
├── flow.py           # FlowStep + Flow + FlowStatus — ordered step definitions
├── registry.py       # FlowRegistry — multi-version flow catalogue
├── executor.py       # FlowExecutor — deterministic, LLM-free runner
├── exceptions.py     # Typed exceptions with traceable context
└── log_utils.py      # Structured per-step logging

Core abstractions

Tool

Tool(
    name="my_tool",
    description="...",
    input_schema=MyInputModel,   # Pydantic BaseModel
    output_schema=MyOutputModel, # Pydantic BaseModel
    fn=my_callable,
)

A tool wraps a plain Python callable together with Pydantic models for strict input/output validation.

FlowStep

FlowStep(
    tool_name="my_tool",
    input_mapping={
        "key_for_tool": "key_from_context",   # flat top-level lookup
        "city": "/user/address/city",         # RFC-6901 pointer into nested context
        "limit": 10,                          # non-string -> literal constant
    },
    output_mapping={"renamed": "value"},      # rename/prune outputs before merge
)

input_mapping maps keys from the accumulated execution context into the tool's input schema. String values are looked up in the context — a plain key is a top-level lookup, and a string starting with / is an RFC-6901 JSON pointer into the nested context (#387) — while non-string values are literal constants.

output_mapping (#386) optionally renames and prunes a tool's outputs before they merge into the context: {context_key: output_key} keeps only the listed output keys, each renamed. Omit it to merge every output key verbatim.

To inject per-request secrets that must never appear in a model-visible schema (auth tokens, account numbers), pass them at execute-time instead of in initial_input:

result = executor.execute_flow(
    "account_overview",
    {"query": "what's my balance?"},          # LLM-visible
    dynamic_params={"billingAccountNumber": "1.60007029"},  # hidden, injected (#316)
)

Flow

Flow(
    name="my_flow",
    version="0.1.0",             # SemVer string; defaults to "0.1.0" if omitted
    description="...",
    steps=[step_a, step_b, step_c],
    deterministic=True,          # metadata annotation; executor is always LLM-free
    trigger_conditions={"intent": "process data"},  # optional metadata
)

An ordered sequence of steps. See AGENTS.md §5 for the full field table (status, tool_schema_hashes, and the input_schema_ref / output_schema_ref string fields with their resolved-property accessors).

A FlowStep runs either a tool (tool_name) or a registered sub-flow (flow_name) — exactly one, never both. Referencing a sub-flow lets you compose reusable flows (issue #75):

fetch_validate = Flow(
    name="fetch_validate",
    description="Fetch and validate.",
    steps=[
        FlowStep(tool_name="fetch", input_mapping={"url": "url"}),
        FlowStep(tool_name="validate", input_mapping={"data": "data"}),
    ],
)
fetch_then_transform = Flow(
    name="fetch_then_transform",
    description="Reuse fetch_validate, then transform.",
    steps=[
        FlowStep(flow_name="fetch_validate", input_mapping={"url": "url"}),  # sub-flow
        FlowStep(tool_name="transform", input_mapping={"data": "data"}),
    ],
)

The executor runs the sub-flow with the step's resolved inputs, merges its output back into the parent context, and attaches the sub-flow's ExecutionResult to the parent StepRecord.sub_result. Sub-flow references are checked for cycles and a configurable max nesting depth (FlowExecutor(max_composition_depth=...), default 10) before execution, raising FlowCompositionError otherwise.

A deadline or CancellationToken passed to execute_flow is forwarded into composed sub-flows, so cancellation and the wall-clock budget are observed at the step boundaries inside a sub-flow — a long sub-flow stops between its own steps rather than only at the parent boundary. The cost report's steps_executed counts the tool invocations a composed step actually drove (recursively), so llm_calls_avoided reflects every tool that ran across the composition.

FlowRegistry

registry = FlowRegistry()
registry.register_flow(flow)
registry.get_flow("my_flow")
registry.list_flows()
registry.match_flow_by_intent("process data")  # basic substring match

An in-memory catalogue of flows.

FlowExecutor

executor = FlowExecutor(registry=registry)
executor.register_tool(tool_a)
result = executor.execute_flow("my_flow", {"key": "value"})

# Version-targeted execution: run an exact registered version instead of the
# latest. Omitting `version` keeps the default (latest) behaviour. The version
# that actually ran is always recorded on `result.flow_version`, so routing,
# audit, and replay can correlate a result with the precise flow definition.
result = executor.execute_flow("my_flow", {"key": "value"}, version="1.2.0")
assert result.flow_version == "1.2.0"

Runs a flow step-by-step with full schema validation and structured logging. No LLM calls are made at any point.

ChainAnalyzer

from chainweaver import ChainAnalyzer, ToolChain

analyzer = ChainAnalyzer(tools=[tool_a, tool_b, tool_c])

# All schema-compatible pairs
matrix: dict[str, list[str]] = analyzer.compatibility_matrix()

# All valid tool sequences up to length 3
chains: list[ToolChain] = analyzer.find_chains(max_depth=3)

# Filter by start or end tool
chains = analyzer.find_chains(max_depth=3, start="tool_a", end="tool_c")

# Promote chains to ready-to-register Flow objects
flows = analyzer.suggest_flows(max_depth=3, min_depth=2)

Discovers schema-compatible tool combinations offline, before any flow is registered or executed. compatibility_matrix() checks that every required input field of a consumer tool appears in the output of the producer with a matching type. suggest_flows() auto-wires input_mapping by name-matching and returns Flow objects ready for FlowRegistry.register_flow().

Data flow

initial_input (dict)
       │
       ▼
 ┌─────────────────────────────────────────────┐
 │  Execution context (cumulative dict)        │
 │                                             │
 │  Step 0: resolve inputs → run tool → merge  │
 │  Step 1: resolve inputs → run tool → merge  │
 │  Step N: resolve inputs → run tool → merge  │
 └─────────────────────────────────────────────┘
       │
       ▼
 ExecutionResult.final_output (merged context)

MCP Integration Concept

ChainWeaver is designed to sit between an MCP server and your agent loop:

MCP Agent
   │  (observes tool call sequence at runtime)
   ▼
ChainWeaver FlowRegistry
   │  (matches pattern → retrieves compiled flow)
   ▼
FlowExecutor
   │  (runs deterministic steps without LLM involvement)
   ▼
MCP Tool Results

In practice:

1. An agent calls tool_a, then tool_b, then tool_c several times with the same routing logic.
2. A higher-level observer detects the pattern and registers a named Flow.
3. On subsequent invocations the executor runs the entire flow in a single call — no intermediate LLM calls required.

ChainWeaver is the library you embed, not the runtime that owns your trace store, auth, side-effect policy, or MCP wiring. Host authors should read docs/runtime-responsibilities.md to see which responsibilities stay on their side of the seam (deciding when to invoke, persisting traces, redacting sensitive outputs, idempotency of side-effect tools, MCP authorisation).

Integrations

ChainWeaver plugs into the MCP ecosystem and the major agent frameworks. Every entry point below ships with a runnable example or recipe.

Install the extra you need: pip install 'chainweaver[mcp]' (or langgraph, openai-agents, langchain, llamaindex). Importing any integration without its extra raises a clear ImportError.

Looking to publish or list ChainWeaver in the MCP registry / awesome-lists / framework directories? See docs/distribution.md.

Error Handling

All errors are typed and traceable:

All exceptions inherit from ChainWeaverError and carry a stable diagnostic code (e.g. CW-E006); the CLI prefixes it on error output and failing StepRecords expose it as error_code. See the full code table in docs/reference/error-table.md.

Standard tool library

chainweaver.contrib.tools ships a curated set of deterministic utility tools so that a new user can compose a meaningful flow on the first afternoon without writing any Tool boilerplate.

from chainweaver.contrib.tools import (
    assert_equal,
    filter_list,
    json_pluck,
    json_set,
    map_list,
    passthrough,
)

The library is deterministic-only: no HTTP, file I/O, database access, RNG, or clocks. Anything stateful belongs in user code. Install with pip install 'chainweaver[contrib]'.

Runnable examples: examples/contrib_pluck_and_set.py, examples/contrib_map_filter.py.

Cost-avoided reporting

Every inter-step transition a naive agent delegates to an LLM is a routing call ChainWeaver eliminates. CostProfile / CostReport turn that into a dollar estimate, and the maintained PROVIDER_PRICES table (dated snapshots, no live HTTP lookup) lets you price it against a real model:

from chainweaver.cost import compute_cost_report

# Build a profile straight from the maintained price table.
report = compute_cost_report(
    steps_executed=6,                 # a six-tool flow
    actual_execution_ms=4.2,
    provider="anthropic",
    model="claude-opus-4-7",
)
print(report)

Cost Avoided Report (estimate)
──────────────────────────────
Steps executed:          6
LLM calls avoided:       5
Est. latency saved:      1500.0ms
Est. cost saved:         $0.1688
Actual execution time:   4.2ms
Priced against:          anthropic/claude-opus-4-7 (as of 2026-05-01)

Every report built from the table carries the snapshot's as_of date so stale prices are visible. Unknown (provider, model) pairs raise CostProfileError rather than guessing. Pass an explicit profile=CostProfile(...) when you have better per-call numbers, or set cost_profile= on FlowExecutor to attach a report to every ExecutionResult. Prices are refreshed by a maintainer-reviewed PR (.github/workflows/update-prices.yml) — never auto-merged.

Export adapters

Hand a compiled flow off to any external agent framework via chainweaver.export:

from chainweaver.export import (
    flow_to_anthropic_tool,
    flow_to_callable,
    flow_to_openai_function,
)

openai_spec = flow_to_openai_function(flow, executor)
anthropic_spec = flow_to_anthropic_tool(flow, executor)
run = flow_to_callable(flow, executor)  # plain dict → dict callable

flow_to_openai_function emits the {"type": "function", "function": {…}} shape OpenAI's chat / responses APIs expect. flow_to_anthropic_tool emits Anthropic's tool_use shape. flow_to_callable wraps the flow as a Callable[[dict], dict] suitable for any framework that accepts arbitrary Python callables.

None of these adapters imports openai or anthropic — they emit dicts and callables only. Runtime integration with those clients is the caller's job.

Runnable example: examples/export_openai_anthropic.py.

Ecosystem bridges (LangChain, LlamaIndex)

chainweaver.integrations.langchain and chainweaver.integrations.llamaindex ship thin bidirectional adapters so existing LangChain BaseTool / LlamaIndex FunctionTool instances can be pulled into ChainWeaver, and ChainWeaver Tool instances can be pushed back out.

from chainweaver.integrations.langchain import (
    from_langchain_tool,
    to_langchain_tool,
)

cw_tool = from_langchain_tool(my_langchain_tool)
lc_tool = to_langchain_tool(my_cw_tool)

Install with pip install 'chainweaver[langchain]' / 'chainweaver[llamaindex]'. Importing either module without the relevant extra raises a clear ImportError.

Plugin discovery

For third-party packages — chainweaver-aws, chainweaver-stripe, … — ChainWeaver follows the same entry-point convention used by pytest, Sphinx, MkDocs, and friends.

Publisher (pyproject.toml):

[project.entry-points."chainweaver.tools"]
aws = "chainweaver_aws:get_tools"

[project.entry-points."chainweaver.flows"]
aws = "chainweaver_aws:get_flows"

Consumer:

from chainweaver import FlowExecutor, FlowRegistry

# Auto-register every tool / flow advertised by an installed plugin.
registry = FlowRegistry(discover_plugins=True)
executor = FlowExecutor(registry=registry, discover_plugins=True)

Discovery is opt-in — importing chainweaver does not trigger plugin imports. Misbehaving plugins (raise on import, return the wrong type) are logged at WARNING and skipped; pass strict=True to discover_tools() / discover_flows() for the loud form.

Runnable example: examples/plugin_discovery.py.

Runtime learning

You don't have to hand-author every flow. ChainWeaver can watch what your agent actually does and propose compiled flows for the repeated, deterministic paths — all offline, with no LLM in the loop.

from chainweaver import ChainObserver, FlowRegistry

observer = ChainObserver()

# Record tool calls as the agent makes them.
observer.record("fetch", {"url": "..."}, {"body": "..."})
observer.record("validate", {"body": "..."}, {"valid": True})
observer.record("transform", {"body": "..."}, {"records": [1, 2, 3]})
observer.end_trace()
# ... many traces later ...

registry = FlowRegistry()
for suggestion in observer.suggest_flows(min_occurrences=3):
    # Suggestions are proposals — review, then promote explicitly.
    print(suggestion.flow.name, suggestion.confidence,
          suggestion.estimated_llm_calls_avoided)
    registry.register_flow(suggestion.flow)

• ChainObserver (#78) mines repeated tool sequences from runtime traces and emits ranked FlowSuggestions — never auto-registered.
• chainweaver record (#226) does the same from a recorded JSONL trace on the command line, writing candidate .flow.yaml files ranked by projected LLM calls avoided. Candidates persist as draft; use chainweaver flows promote ... --to reviewed, then --to active. Ignored candidates remain suppressed in later runs unless --include-ignored is passed.
• ChainWeaverService (#101) ties the observer, the static ChainAnalyzer, and an optional offline LLM proposer into a continuous analyze → propose → govern → promote loop with an in-process governance gate and adoption metrics.

Runnable example: examples/chain_observer.py.

Roadmap

Milestones below mirror the GitHub milestones; see CHANGELOG.md for a per-release feature breakdown.

Curious how ChainWeaver compares to LangChain, LangGraph, Prefect, Dagster, or Temporal? See docs/comparisons.md.

Command-line interface

ChainWeaver ships a chainweaver console script with the following subcommands. Reading .flow.yaml files needs the YAML extra (pip install 'chainweaver[yaml]' — also listed in Installation). The run example below uses a flow shipped under examples/, so it should be invoked from the repository root.

# Run a flow from disk — no Python required.
chainweaver run examples/double_add_format.flow.yaml \
    --tools examples.simple_linear_flow \
    --input '{"number": 5}'

# Serve a flow as MCP tools (needs chainweaver[mcp]) — agents call the whole
# compiled flow as one deterministic tool. See docs/mcp-server.md.
chainweaver serve examples/double_add_format.flow.yaml \
    --tools examples.simple_linear_flow

# Validate a flow file (used by CI gates and editor tooling).
chainweaver validate flows/etl.flow.yaml
chainweaver check flows/                  # whole-directory variant

# Scaffold a runnable first flow project (tools + flow file + run script).
chainweaver init my-first-flow --template linear --with-tests

# Render a flow as ASCII, Graphviz DOT, or Mermaid. Discover it from a directory
# of flow files, an installed package's entry points, or the default registry.
chainweaver viz my_flow --discover-dir flows/ --format dot | dot -Tpng -o my_flow.png
chainweaver viz my_flow --discover-dir flows/ --format mermaid
chainweaver viz --result trace.json --format mermaid   # overlay a real run

# Explain a flow deterministically (LLM-free) for review — paste into a PR.
chainweaver explain my_flow --discover-dir flows/ > flow-review.md

# Inspect a flow's structure (table or JSON). `flows list` previews what is
# discoverable so you can see what `inspect`/`viz` can target.
chainweaver inspect my_flow --discover-dir flows/ --format json
chainweaver flows list --discover-dir flows/

# Check that your environment is ready before running anything.
chainweaver doctor flow --profile first-run

# Inspect a coding-agent workspace's MCP / observe setup (read-only).
chainweaver doctor vscode --workspace .

# Install tab-completion for your shell (bash/zsh/fish).
chainweaver --install-completion

# Analyze ExecutionResult traces — bottlenecks, p50/p95/p99 across runs,
# and per-step / per-tool retry / skip / fallback / failure aggregates.
chainweaver profile trace_a.json trace_b.json --format json

# Compare two ExecutionResult JSON files step-by-step.
chainweaver diff baseline.json current.json --perf-tolerance 25

# Observed-determinism attestation: run N inputs × M repeats.
chainweaver attest flows/etl.flow.yaml --tools my_pkg.tools --runs 50 --repeats 3

# Advisory optimization suggestions for a saved flow.
chainweaver suggest flows/etl.flow.yaml --tools my_pkg.tools --trace trace_a.json

# Mine candidate flows from a recorded JSONL tool trace (offline, no LLM).
chainweaver record examples/agent_tool_trace.jsonl --output-dir candidates/
chainweaver flows promote candidates/suggested__fetch__validate.flow.yaml --to reviewed
chainweaver flows promote candidates/suggested__fetch__validate.flow.yaml --to active

# Run one continuous-analysis service pass and report flow proposals.
chainweaver service --tools my_pkg.tools --trace trace.jsonl

# Check saved flows for tool schema drift against the live registry.
chainweaver doctor flow flows/ --check-drift --tools my_pkg.tools

# Property-based fuzzing: generate cases, check invariants, save/minimize failures.
chainweaver fuzz flows/etl.flow.yaml --tools my_pkg.tools \
  --property my_pkg.props:no_unauthorized_action --runs 1000 --seed 42 \
  --minimize --save-failures failures/

run is the fastest path from a fresh install to seeing a flow execute: point it at a .flow.yaml/.flow.json file, pass --tools <module> (the import path of a Python module that exposes Tool instances at top level), and supply the initial input as JSON. Hand-authored flow files must declare a type: Flow (or type: DAGFlow) discriminator at the top — see the flow file format reference. Most reporting subcommands also accept --format json for machine consumption (inspect, validate, check, run, profile, diff, attest, suggest, doctor); the exceptions are viz, which uses --format ascii|dot|mermaid, explain, which uses --format md|text, and dump-schema, which writes a raw JSON Schema and has no --format flag. The result-producing commands (inspect, validate, check, profile, diff, attest) wrap their --format json output in a stable, versioned envelope ({"schema_version", "status", "data", "errors"}) so automation can branch on status / error codes — see machine-readable output. All subcommands share the same exit-code contract (0 success, 1 business-logic error, 2 file-not-found / argument error), and the CLI ships tab-completion (chainweaver --install-completion).

inspect and viz resolve flows from disk or a registry. Pass --file <path>, --discover-dir <dir>, or --discover-entry-points to resolve a flow without writing any Python (issue #381); chainweaver flows list previews what is discoverable. With no discovery flag they fall back to a process-scoped, in-memory registry installed programmatically — running chainweaver inspect my_flow with neither a flag nor a configured registry exits 1 with No registry configured. Call chainweaver.cli.set_default_registry(...) before invoking the CLI.. To wire the default-registry path, use a small entry script:

# my_cli_entry.py
from chainweaver import FlowRegistry
from chainweaver.cli import main, set_default_registry
from my_app import build_registry  # returns a populated FlowRegistry

set_default_registry(build_registry())
main()

See docs/cli.md § Programmatic registration for the full pattern, including why the split exists (file-oriented commands stay zero-config, registry-oriented commands stay introspection-friendly).

Development

New contributors: see Your first contribution in CONTRIBUTING.md for the good-first-issue / good-first-ai-issue onramp and the step-by-step path to your first PR.

# Install with dev dependencies
pip install -e ".[dev]"

# Run tests
python -m pytest tests/ -v

# Run the examples
python examples/simple_linear_flow.py   # simple arithmetic flow
python examples/etl_flow.py             # ETL flow
python examples/mcp_search_flow.py      # MCP-style search & summarize flow
python examples/naive_vs_compiled.py    # naive vs compiled timing comparison
python examples/coding_agent_pr_review.py
python examples/coding_agent_changelog.py
python examples/coding_agent_debug_log.py

License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.