An alpha project from HOL.orgAlpha · Python SDK on PyPI

Give every agent a governed decision environment.

CIGAR turns source material into provenance-bearing context, carries only bounded authority across agent handoffs, treats external actions as recoverable effects, and preserves evidence for audit and replay.

CIGARContext Intelligence Graph Agentic RuntimeCIGAR is a model-agnostic runtime for governed context, bounded agent authority, recoverable actions, and replayable evidence.
Explore CIGAR HoneyRead the architecture

PUBLISHED PYTHON SDK

Install hol-cigar from PyPI.

The developer-preview SDK is available for Python 3.14. The distribution name is hol-cigar; application code imports the cigar_sdk namespace. The full Honey desktop archive remains a separate release.

python3.14 -m pip install "hol-cigar==0.9.1"
python3.14 -c 'import cigar_sdk; print(cigar_sdk.__version__)'
Open on PyPIView source on GitHubRead SDK documentation

THE SYSTEM AROUND THE PROMPT

The prompt is only part of the system.

Agents assemble context, call tools, hand work to other agents, and retry after partial failures. Without an explicit runtime contract, source text can blur into instruction authority, prompt construction becomes invisible, ambiguous tool outcomes look final, and audit logs omit the context that shaped a decision.

01
Opaque prompt assembly

Deterministic context bundles with manifests and provenance

02
Authority expands during delegation

Recipient-bound handoffs with attenuated capabilities and budgets

03
Tool uncertainty is flattened into success or failure

Durable effect intent, authorization, observation, reconciliation, and compensation

04
Logs cannot recreate the decision environment

Evidence-bound replay without re-running live side effects

FOUR RUNTIME CONTRACTS

Context that can explain itself. Authority that stays bounded.

Governed context

Observe filesystem and Git sources, apply policy before disclosure, enforce budgets, and compile immutable bundles with stable identities.

Bounded coordination

Fork private work, create signed recipient-bound handoffs, attenuate authority, record typed results, and merge against an exact base.

Recoverable effects

Separate intent from dispatch. Preserve UNKNOWN after ambiguous execution and require explicit reconciliation or compensation.

Replayable evidence

Bind inputs, policy, runtime fingerprints, outputs, effects, observations, and uncertainty without collecting hidden model reasoning.

HOW IT WORKS

A control and evidence layer between sources and agents.

CIGAR is model-agnostic. Applications keep their chosen models and orchestration logic while CIGAR provides a stable contract for context, delegated authority, effects, and evidence.

CIGAR architecture: ingest sources, apply context contract and policy, compile context for agents, govern handoffs and effects, and replay evidence
CIGAR ingests immutable source snapshots, applies a purpose, rules, and budget contract, retrieves authorized context, compiles a bundle and manifest, and delivers bounded context to the agent or model. Its runtime also governs scoped handoffs and intent-first effects, while provenance, authorization, traceability, and evidence verification support replay.

WORKFLOWS

One runtime contract across the agent workflow.

  1. 01Context compilationSelect, deduplicate, budget, and explain context before model use.
  2. 02Multi-agent handoffGive a recipient the work and authority it needs—no more.
  3. 03Protected disclosureEnforce purpose, principal, project, and trust rules before content is revealed.
  4. 04External action recoveryRetain an honest UNKNOWN state when execution is ambiguous.
  5. 05Audit and replayReconstruct declared inputs and replay recorded observations without live provider or effect egress.

What CIGAR is not

CIGAR is not a foundation model, chatbot, autonomous scheduler, hosted agent platform, or a substitute for application orchestration. It is the runtime contract around context, authority, effects, and evidence.

Available for evaluation · public release pending

Meet CIGAR Honey 0.9.1 alpha.

Honey is the first bounded CIGAR profile for local, hands-on evaluation on Apple-silicon macOS. It packages the CLI, local daemon, MCP and Claude Code integrations, language SDKs, deterministic demos, and the context, handoff, effect, and replay workflows behind them.

See what is in Honey

FAQ

CIGAR, in plain terms.

Is CIGAR a model?

No. CIGAR is model-agnostic infrastructure around context, delegated authority, effects, and evidence.

Does CIGAR store chain-of-thought?

No. CIGAR retains observable inputs, outputs, policy decisions, effects, and uncertainty without requesting hidden model reasoning.

Does CIGAR replace an orchestrator?

No. The application keeps its orchestration logic. CIGAR supplies governed runtime contracts the orchestrator can use.

Can CIGAR support more than one agent?

Yes. The protocol includes context spaces, checkpoints, recipient-bound handoffs, typed results, and exact-base merge semantics. Honey demonstrates a bounded two-agent local profile.