tuliplabs-ai

tulip-agents: Gate Consequential Actions in AI Agents

Tulip Agents is a Python agent framework built around an admission gate that checks consequential actions against policy before they run. It is designed for teams building agents or adding enforceable controls, human approval, and tamper-evident audit records to existing ones.

ArchivedPythonApache-2.0AI AgentsAgent FrameworkSecurity
Stars
1
Forks
1
Last commit
24 days ago
Contributors
3
Releases, 12 months
48

Our take

Archivedarchived by its owners on GitHub, read-only

Health score

50/100

How it is scored
Activity
85
Community
13
Issues
21
Pull requests
82
  • Merges more pull requests than 92% of the projects we track
  • Releases more often than 86% of the projects we track
  • Fewer contributors than 82% of the projects we track

Tulip's design makes action control a runtime boundary rather than a prompt convention, but its archived status and nearly single-contributor development make it a risky foundation for new production adoption without a maintenance plan.

Good fit if

  • You need to inspect or adapt a Python implementation of policy-gated agent actions, and can maintain a fork because the repository is archived.
  • Your team wants human approval and tamper-evident records around agent tool calls, and understands that the audit chain is not notarized.
  • You are evaluating a broad agent SDK and can validate the optional integrations and backends you intend to use.

Look elsewhere if

  • You need an actively maintained upstream project to take responsibility for fixes and ongoing releases.
  • Your team cannot absorb the risk of development being concentrated in one contributor.
  • You need a simple agent library without runtime governance, multi-agent features, or the additional integration surface.
All health signals
Last commit2026-10-23 (24 days ago)
Commits, last 90 days100+
Releases, last 12 months48 (latest v2.28.0, 2026-10-10)
Contributors3 (top contributor: 100% of commits)
Issues closed, last 90 days0+
Pull requests merged, last 90 days88 (typically merged in 0 days)
Project age2 months

Checked on 2026-11-16 with the GitHub API.

Overview

Tulip Agents is a Python agent framework built around a structural admission gate: before an agent can perform a consequential action, runtime code checks it against policy and can allow, hold for human approval, or deny it. The model does not control or bypass that gate. Decisions are recorded in a tamper-evident audit chain, addressing the risk that prompt-based rules alone may not stop an agent from performing an unsafe action.

The project can also serve as a full SDK for building agents, with multi-agent workflows, grounding, memory, retrieval, observability, and a server. Its framing is broader than a standalone safety wrapper: teams can use the admission controls around an existing agent or build a governed agent with Tulip. The repository is archived, however, so it is read-only despite its recent commits and releases.

Key Features

  • admit() and gate_tool() place policy checks before side effects, with decisions to allow, hold, or deny.
  • Human approvals can pause a run and resume an approved call using its original arguments.
  • The audit trail links records into a SHA-256 chain that can be checked for tampering; it is not notarized.
  • Idempotent tools can refuse duplicate calls with the same name and arguments, including across retries.
  • The SDK provides agent loops, multi-agent patterns, a cognitive router, and cross-process A2A support.
  • GSAR grounding classifies claims against evidence and can regenerate, replan, or abstain below a threshold.
  • Optional integrations include MCP, RAG stores, checkpoint backends, observability, evaluation, and a FastAPI server.

Use Cases

  • A team adding an agent to a payment or customer-support workflow can gate refunds or record changes and require a person to approve production actions.
  • Platform engineers can wrap existing agent tools so policy checks and audit records apply before deployments, deletions, or other consequential operations.
  • Teams building multi-step or multi-agent workflows can use the SDK's pipelines and router while keeping action admission in the runtime.
  • Developers testing agent safeguards can run the bundled rogue-agent challenge offline, or connect a model endpoint to test their own setup.

What you need

Detected in the repository

  • Python >=3.11 (from pyproject.toml)
  • A Dockerfile, so it can run in a container
  • A test suite and automated checks on GitHub Actions

License in plain words

Apache-2.0permissive

  • Commercial use: yes
  • Modify and redistribute: yes
  • You must keep: the license, the NOTICE file and a note of your changes
  • Share your changes: no
  • Includes an explicit patent grant from the contributors:

A summary, not legal advice: the LICENSE file is what applies.

Getting Started

Install the package and try the offline challenge:

pip install tulip-agents
python -m tulip.rogue

For installation extras, agent examples, and configuration, see the README and documentation.

Alternatives

  • agent-governance-toolkit: It provides broader governance controls and integrations for existing agent frameworks, while Tulip centers on a policy gate for consequential actions.
ProjectLanguageLicenseStarsStatus
tulip-agentsPythonApache-2.01Archived
agent-governance-toolkitPythonMIT6.6kActive

Considerations

  • GitHub marks the repository archived and read-only. The measured history shows recent activity and releases before archival, but that does not imply future maintenance.
  • Most contributions come from one person, and no issues were closed in the measured recent period. Factor in the likelihood that your team may need to diagnose and maintain changes itself.
  • The runtime requires Python 3.11 or newer. Docker is supported; particular integrations and backends are optional extras, so check their requirements before selecting them.
  • The audit trail is described as tamper-evident, not notarized. It should not be treated as legally authoritative without additional signing.
  • The project is licensed under Apache-2.0. Review the license and notices when distributing or modifying it.

Found this useful?

Share it with someone who would like tulip-agents.