TokenOps

Run-aware token governance for multi-agent systems.
Cap spend and steer behavior across a whole agent workflow — not per request — with a shared ledger and in-path enforcement.

_{Governed Research → Summarize run: the budget cap halts spend mid-run, then the Dashboard attributes cost per agent. Full video.}

TokenOps is a control plane + SDK for agent stacks. Entry agents register a run; every LLM and tool crossing shares one run_id and one ledger. Policies can halt, mutate, or inject before the next call executes — so a research → summarize → review pipeline stays inside a single budget even across processes.

Why · Architecture · Install · Quick start · Onboarding · Demos · Comparison · Make targets · Roadmap

Why TokenOps

• Govern the run, not the request. One run_id spans every model, tool, and A2A hop in a workflow.
• Shared ledger across processes. Spend, inflight, and halt live in SQLite so multi-agent stacks cannot each burn the full cap locally.
• In-path enforcement. wrap_complete runs detect → decide → apply before the next LLM call; Chronicle @boundary + a crossing hook ingest tool spend.
• Steer or stop. Actuators: HALT · MUTATE · INJECT · reject/queue — not just post-hoc analytics.
• Batteries included. Control plane (:7700), Admin + Dashboard UI, ten seeded policies, and runnable A2A benches (two-agent, triad, LangChain brief).

Architecture

TokenOps is two layers that share one artifact, the run: a control plane that registers runs and stores budgets/policies, and an in-process SDK that enforces at every boundary crossing.

flowchart LR
    subgraph PLANE["Control plane (:7700)"]
        R["POST /v1/runs"] --> DB[("SQLite TOKENOPS_DB<br/>registrations · budgets · policies · ledger")]
        UI["Admin + Dashboard"] --> DB
    end

    subgraph AGENTS["Agent processes (SDK)"]
        E["Entry agent<br/>tokenops_run"] -->|"register_run"| R
        E -->|"X-TokenOps-Run-Id"| D["Downstream agents<br/>tokenops_run"]
        E & D -->|"wrap_complete"| G["Governor<br/>pre_call → detect → decide → apply"]
        E & D -->|"@boundary + crossing hook"| G
        G --> L["Shared ledger<br/>(same run_id)"]
    end

    L --> DB
    DB -->|"governance_config_for"| G

Chronicle records decision boundaries; TokenOps attaches as the cost/governance observer on live crossings. See Chronicle for record-and-replay.

Install

pip install agent-tokenops

# With example / bench extras (LangChain, ddgs):
pip install "agent-tokenops[examples]"

# From source (development):
pip install -e ".[dev,examples]"

Prerequisites: Python 3.10+; agent-tokenops; either a running control plane (TOKENOPS_URL + shared TOKENOPS_DB) or TOKENOPS_EMBEDDED=1 for single-process / tests. LLM API keys only for real model calls. FastAPI only if you use instrument_app.

PyPI name is agent-tokenops; import is still tokenops (same pattern as Chronicle). See RELEASING.md for releases.

Quick start

make install
cp .env.example .env   # optional API keys for demos / your agents

make db-reset          # optional: clean SQLite + seed governance from default.yaml
make run               # control plane :7700 + Admin/Dashboard :8501

Wire governance into an agent: instrument_app once, then tokenops_run per request. The UI sends task only; intent / mode come from agent config on instrument_app.

from tokenops import ControlPlaneClient, instrument_app, tokenops_run
from tokenops.control import wrap_complete, with_governance_errors
from tokenops.providers import complete

client = ControlPlaneClient.from_env()  # TOKENOPS_URL or embedded Store

async def handler(payload: dict, headers: Mapping[str, str]) -> dict:
    with tokenops_run(client=client) as bound:
        governed = wrap_complete(
            bound.governor, bound.controls, bound.attr,
            provider=provider, model=model,
            dispatch=complete, service="planner",
        )
        run_agent(..., complete_fn=governed)

app = create_a2a_app(..., handler=with_governance_errors(handler))
instrument_app(app, service="planner", intent="triad_plan",
               provider=provider, model=model)

Non-FastAPI: TokenOps does not yet ship middleware for other frameworks. Use bind_request_context(RequestContext(headers=..., payload=..., service=...)) then with tokenops_run():, or pass those kwargs explicitly to tokenops_run.

Point agents at the plane and share one DB:

export TOKENOPS_URL=http://localhost:7700
export TOKENOPS_DB=tokenops.db   # plane + all agents
make control-plane               # :7700
make ui                          # Admin + Dashboard :8501

New here? Onboarding guide (prereqs, bare-min integrate, FAQ, current limits).

Full integration checklist: .cursor/skills/integrate-tokenops/SKILL.md · triad deep dive: docs/guides/field-guide-add-tokenops.md.

Demos

Each bench is a multi-agent stack with a shared run ledger. Start the plane, agents, and Admin UI with one make target.

make demo              # plane + research/summarize + Admin UI
make demo-triad        # plane + planner/researcher/writer
make demo-brief        # plane + scout/analyst/editor

Docker:

docker compose up --build
# optional UI: docker compose --profile ui up --build
# two-agent stack:
docker compose -f docker-compose.examples.yml up --build

See examples/README.md and docs/control-plane-deploy.md.

How TokenOps compares

TokenOps is not a gateway or a tracing dashboard. It governs the run — a full agent workflow — and sits alongside the tools you already use for routing and observability.

What this does not do: replace your LLM gateway, replace Chronicle-style record-and-replay, or host a SaaS control plane for you. Fail-closed integrity (refuse on missing registration) is optional and upcoming.

Longer table with logos: docs/product/comparison.md.

Make targets

Environment variables

Production / multi-process: set TOKENOPS_URL; agents must not mount /v1/runs. Tests: TOKENOPS_EMBEDDED=1 (or omit URL).

Project structure

Only src/tokenops/ is the installable package. Demos and benches stay under examples/.

src/tokenops/              # installable package
├── server/                # control plane (:7700, POST /v1/runs)
├── control/               # SDK: ledger, policies, wrap_complete, crossing hook
├── providers/             # OpenAI / Anthropic complete dispatch
├── config/                # default.yaml governance seed
└── ui/                    # Admin + Dashboard (Streamlit)
examples/                  # A2A benches (two-agent, triad, brief) + Chat/Simulator
benchmarking/              # MetaGPT / browser-use live harness
docs/                      # architecture, policies, guides, product
tests/                     # unit + e2e

Roadmap

TokenOps is early (0.x). Near-term:

• User/tag segment-scoped budgets (machinery exists; seed is run-only today).
• Optional fail-closed mode on missing registration or exceeded budget.
• Remote observe / decide (fatter plane) for multi-host stacks.
• Documentation site.

Status of each control-plane job: docs/control-plane-status.md. Ideas welcome via GitHub issues.

Documentation

• Onboarding — prereqs, bare-min integrate, FAQ, current limits
• Field guide — triad deep dive + screenshots
• Control plane status
• Architecture
• Run attribution
• Control plane deploy
• Examples
• Product: comparison · shared ledger

Contributing

Issues and PRs are welcome. See CONTRIBUTING.md, CODE_OF_CONDUCT.md, and SECURITY.md.

make install
make lint
make test

Contributors

Thanks to everyone who has contributed.

If TokenOps saves you a runaway agent bill, please ⭐ star the repo so more people can find it.