coordpy

A Python-first agent development kit (ADK) for building LLM agents and agent teams with a familiar Agent / Tool / Runner / Session / State / Memory / Artifacts mental model — that gives you content-addressed capsule audit, provenance, and replay for free underneath. You write import-and-code ADK; every model call, tool call, and handoff automatically seals into a hash-chained capsule you can re-verify from bytes and replay against another model. The CLI is a secondary runtime surface.

PyPI: coordpy-ai · import: coordpy · library front door: coordpy.adk

Why CoordPy

Multi-agent stacks usually pass context around as raw prompts and JSON. That scales until the prompt grows past the model's useful window and the run silently devolves into token cramming — and then when something breaks you can't reconstruct what each agent saw.

CoordPy attacks both edges of that frontier:

• Bounded context, not token cramming. Each agent sees the team instructions plus the latest N visible handoffs (default N=4), not the full transcript. Every team run reports the bounded-context savings in real tokens, not vibes.
• Auditable handoffs. Every agent output seals into a TEAM_HANDOFF capsule with a content-derived ID, declared parents, a per-handoff byte/token budget, and witness fields (prompt_sha256, model_tag) that prove which prompt produced which output.
• Replayable runs. A team_result.json manifest records each turn's prompt, generation params, and capsule CID. coordpy-team replay re-runs the same prompts on a different backend/model at the original sampling settings, so the audit story holds across models.

Install

Requires Python 3.10 or newer.

pip install coordpy-ai

Verify:

coordpy --version           # coordpy 1.2.1 (coordpy.sdk.v3.43)
python -c "import coordpy; print(coordpy.__version__)"

The first parenthetical (coordpy.sdk.v3.43) is the research-line tag exposed at coordpy.SDK_VERSION. It tracks the underlying research programme and is independent of the PyPI version.

The only required dependency is NumPy. Optional extras:

Quickstart — your first agent (Python)

The front door is import-and-code. Tools are plain typed functions (a trailing tool_context is auto-injected and hidden from the model); a turn is a stream of events; the answer is the one that satisfies is_final_response():

from coordpy.adk import Agent, InMemoryRunner, ToolContext
from coordpy.llm_backend import backend_from_env

def lookup_population(city: str, tool_context: ToolContext) -> dict:
    """Return the population of a city."""
    table = {"tokyo": 37_400_000, "paris": 11_100_000}
    tool_context.state["last_city"] = city
    pop = table.get(city.lower())
    return ({"status": "success", "city": city, "population": pop}
            if pop else {"status": "error", "error": f"unknown city {city!r}"})

assistant = Agent(
    name="geo_assistant",
    model=backend_from_env(),          # any Ollama / OpenAI-compatible backend
    instruction="Answer population questions; use the lookup tool.",
    tools=[lookup_population],
    output_key="answer",
)

runner = InMemoryRunner(agent=assistant, app_name="quickstart")
runner.session_service.create_session(
    app_name="quickstart", user_id="u1", session_id="s1")

for event in runner.run(user_id="u1", session_id="s1",
                        new_message="How big is Tokyo?"):
    if event.is_final_response():
        print(event.text)

Set a backend with the COORDPY_* environment variables (local Ollama or any OpenAI-compatible provider; see the env-var blocks below). For a no-network version that needs no model, run examples/adk_quickstart.py; for a small multi-agent app (researcher → writer, with tools + artifacts + memory), run python -m coordpy.adk.examples.research_assistant (source).

Capsule audit, provenance & replay — for free

You wrote plain ADK code. Underneath, every step sealed into a content-addressed, hash-chained capsule trail you can verify from bytes:

view = runner.session_capsule_view("s1")   # the coordpy.capsule_view.v1 chain
assert runner.verify_session("s1")          # re-verified from bytes; tamper-evident
print(runner.session_root_cid("s1"))        # deterministic for identical inputs

That is CoordPy's distinctive layer: an ADK that is auditable and replayable by construction. coordpy.adk.Agent is the primary surface; the higher-level preset teams and the CLI below build on the same capsule machinery. (coordpy.adk.Agent is additive and separate from the legacy coordpy.Agent / create_team surface, which is unchanged.)

Command-line interface (secondary)

The CLI is a secondary runtime surface. The bundled coordpy-team command drives a curated preset against a local Ollama endpoint or any OpenAI-compatible provider:

# Local Ollama (no API key needed)
export COORDPY_BACKEND=ollama
export COORDPY_MODEL=qwen2.5:14b
export COORDPY_OLLAMA_URL=http://localhost:11434

coordpy-team run \
    --preset quant_desk \
    --task examples/scenario_bullish.txt \
    --out-dir /tmp/desk-run

That writes four files into /tmp/desk-run:

Re-verify the chain from bytes alone:

coordpy-capsule verify-view --view /tmp/desk-run/team_capsule_view.json

Replay the same prompts on a different model and compare:

coordpy-team compare \
    --preset quant_desk \
    --task examples/scenario_bullish.txt \
    --backend ollama --model qwen2.5:14b \
    --replay-backend ollama --replay-model gemma2:9b \
    --out-dir /tmp/desk-compare

The compare report shows whether the per-turn prompt SHAs match and whether the synthesizer's parsed ACTION agrees across models.

Structured profile runs (RunSpec → RunReport)

The original structured-research path is still stable and useful for reproducible, profile-driven evaluation runs:

import coordpy

report = coordpy.run(coordpy.RunSpec(
    profile="local_smoke",
    out_dir="/tmp/cp-smoke",
))

assert report["readiness"]["ready"]
assert report["provenance"]["schema"] == "coordpy.provenance.v1"
assert report["capsules"]["chain_ok"]

print(report["capsules"]["root_cid"])

coordpy.run writes seven files into out_dir. The two you will reach for most are product_report.json (the same shape as the returned dict) and capsule_view.json (the sealed capsule chain that coordpy-capsule verify re-hashes); the others (provenance.json, meta_manifest.json, readiness_verdict.json, product_summary.txt, sweep_result.json) are always written and are useful for audit. The root_cid is the SHA-256 of the run's RUN_REPORT capsule; it is stable for a given input but differs between runs because provenance includes a wall-clock timestamp.

Console scripts

New to this repo? coordpy-subject is the fastest way to learn what is stable (the SDK/CLI), what is canonical-experimental (the architecture north-star lineage + the most recent research chain), and what is historical — without reading the milestone archive. See docs/reference/W144_COORDPY_SUBJECT_REGISTRY.json.

The research-profile chain is still useful for the structured RunSpec → RunReport path:

coordpy --profile local_smoke --out-dir /tmp/cp-smoke
coordpy-ci --report /tmp/cp-smoke/product_report.json --min-pass-at-1 1.0
coordpy-capsule view   --report /tmp/cp-smoke/product_report.json
coordpy-capsule verify --report /tmp/cp-smoke/product_report.json

To exercise coordpy-import against the bundled mini fixture (no external file required):

FIXTURE=$(python -c 'import coordpy, os; print(os.path.join(os.path.dirname(coordpy.__file__), "_internal/tasks/data/swe_real_shape_mini.jsonl"))')
coordpy-import --jsonl "$FIXTURE" --out /tmp/audit.json

Agent teams in Python

AgentTeam.from_env reads its backend from COORDPY_* environment variables and requires a configured backend to run — either a reachable Ollama server or an OpenAI-compatible API key. To run a team without a network, see the SyntheticLLMClient example below.

from coordpy import AgentTeam, agent

team = AgentTeam.from_env(
    [
        agent("planner",    "Break the task into 2-3 concrete steps."),
        agent("researcher", "Gather the facts that matter."),
        agent("writer",     "Write the final answer for the user."),
    ],
    model="gpt-4o-mini",
    backend_name="openai",
    team_instructions=(
        "Reuse visible handoffs instead of restating the task."
    ),
    max_visible_handoffs=3,
    task_summary="Answer briefly using only the prior handoffs.",
)
result = team.run("Explain what coordpy does.")
print(result.final_output)

# Bounded-context savings vs naive token cramming.
print(result.cramming_estimate())

# Dump a four-file replayable bundle.
result.dump("/tmp/team-run")

For curated multi-role teams (quant desk, code review, research writer) skip the role-prompt typing and use the bundled coordpy.presets:

from coordpy import presets

team = presets.quant_desk_team()
result = team.run(open("examples/scenario_bullish.txt").read())

Local Ollama:

export COORDPY_BACKEND=ollama
export COORDPY_MODEL=qwen2.5:0.5b
export COORDPY_OLLAMA_URL=http://localhost:11434

OpenAI-compatible provider:

export COORDPY_BACKEND=openai
export COORDPY_MODEL=gpt-4o-mini
export COORDPY_API_KEY=...
# Optional, for non-default providers:
# export COORDPY_API_BASE_URL=https://your-provider.example/v1

To run a team without a network or an API key, pass a SyntheticLLMClient directly:

from coordpy import create_team, agent
from coordpy.synthetic_llm import SyntheticLLMClient

team = create_team(
    [agent("planner", "..."), agent("writer", "...")],
    backend=SyntheticLLMClient(default_response="ok"),
)
print(team.run("hi").final_output)

The bundled examples/ ladder (01_quickstart.py, 02_quant_desk.py, 03_replay_and_audit.py) drives the same public surface end-to-end against a real backend.

Public surface

The experimental surface ships in the same wheel for reproducibility and audit. Pin against coordpy.__experimental__ if you depend on it.

Limitations

• coordpy works at the capsule layer. It does not provide transformer-internal trust transfer or hidden-state access.
• The bundled cross-host evidence comes from the small two-node lab where it was generated. Behaviour at larger scales has not been measured.
• Not peer-reviewed. The code, tests, results notes, and theorem registry are public so they can be challenged.

Where to go next

• Contributing: CONTRIBUTING.md
• Releasing to PyPI: RELEASING.md
• Security policy: SECURITY.md
• Changelog: CHANGELOG.md

License

MIT. See LICENSE.