OpenGAP: an agent defined as a git repository, where only the identity ports and the runtime stays behind
A framework-agnostic, git-native standard for defining AI agents
At a glance
- What is it?
- OpenGAP, the Git Agent Protocol, defines an AI agent as a set of files in a git repo, with agent.yaml and SOUL.md required. The opengap CLI validates it, but orchestration, tool execution and memory I/O are explicitly not part of what travels.
- Who is it for?
- Use OpenGAP if you want agent prompts, rules and roles kept in git with diffs, branches and a validator, and you accept that the framework still runs the agent. Skip it if you expected a runtime or a finished adapter for your framework, since the README says runtime orchestration stays where it is.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 136 days ago.
- What is it written in?
- Mainly TypeScript, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.
Star history
- open-gitagent/opengap3,099 ★ · +168
| Repository | 2026-09-12 | 2026-11-15 |
|---|---|---|
| open-gitagent/opengap | 2.9k | 3.1k |
168 Stars in the last 64 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
Two required files make a repository an agent, everything else is opt-in
The standard is a file layout. Drop files into any git repo and it becomes a portable agent definition. Only two are required: agent.yaml, the manifest, and SOUL.md, the identity. The manifest carries name, version, model, skills, tools and compliance. SOUL.md covers identity, personality, communication style and values.
Optional files add behaviour. RULES.md holds hard constraints, must-always and must-never rules and safety boundaries. DUTIES.md holds segregation of duties policy and role boundaries. AGENTS.md is a framework-agnostic fallback set of instructions. The README tells you to add what you need and ignore the rest.
The consequence of having so small a required core is that two OpenGAP agents can look very different, and a repo that only has the two files conveys little about tools or behaviour. The format standardises where things go, not how much you say. The full layout continues with capability folders such as skills/, tools/, knowledge/, memory/, hooks/ and workflows/ that the patterns refer to.
The CLI has had three names, and the old command still installs as an alias
The package on npm is @open-gitagent/opengap. It was previously published as @open-gitagent/gapman and originally as @open-gitagent/gitagent. The package.json defines two binaries pointing at the same file, opengap and gitagent, so the gitagent command is still installed as an alias for backward compatibility.
The release titles show the path. v0.3.2 is described as publishing gapman both scoped and unscoped, v0.4.0 as OpenGAP, a rebrand from gapman, on 2026-05-24, and v0.5.0 on 2026-07-02 as Claude session continuation with --continue, --resume and --session-id.
For a reader this means older tutorials, scripts and CI files may call a package or command under a previous name. The README itself still uses gitagent in several places, including the sentence saying gitagent fixes the lack of a portable agent definition. When you search for help, expect all three names. The CLI needs Node 18 or later according to the engines field, and the repository is MIT licensed.
Segregation of duties is validated, not enforced at runtime
The most distinctive pattern is segregation of duties. You define roles such as maker, checker, executor and auditor, a conflict matrix saying which roles cannot be the same agent, and handoff workflows, all in agent.yaml and DUTIES.md. The README says the validator catches violations before deployment. A trimmed example from it:
compliance:
segregation_of_duties:
roles:
- id: maker
description: Creates proposals
permissions: [create, submit]
- id: checker
description: Reviews and approves
permissions: [review, approve, reject]
conflicts:
- [maker, checker] # maker cannot approve own work
assignments:
loan-originator: [maker]
credit-reviewer: [checker]
handoffs:
- action: credit_decision
required_roles: [maker, checker]
approval_required: true
enforcement: strictThe word to notice is validator. The check happens when you run opengap validate, which the CI pattern suggests running on every push. Whether a running system honours the policy depends on the framework executing the agent, so strict enforcement here is a statement in the definition, and a deployment that ignores the file is not stopped by it. The compliance language names FINRA, the Federal Reserve and the SEC.
What ports cleanly and what stays in the framework
The porting section is the project's own statement of its limits. Agents built in frameworks such as NVIDIA AIQ, LangGraph or CrewAI have identity split across config files, Jinja2 templates and Python code, and OpenGAP extracts the identity layer into a portable, versionable format.
What ports cleanly is system prompts, persona definitions, hard constraints, tool schemas, role and segregation of duties policies, and model preferences. What stays in the framework is runtime orchestration such as state machines and graph wiring, live tool execution, memory I/O and iterative loops.
The practical result is that adopting OpenGAP does not let you drop your framework, and a LangGraph graph or a CrewAI crew keeps being the thing that runs. You gain a reviewable, diffable record of the agent's character and constraints. If you were hoping for a standard that also captures control flow, that part is not covered. The repository includes an example based on NVIDIA's AIQ Deep Researcher, under examples/nvidia-deep-researcher.
SkillsFlow fixes the order of steps in YAML, with no model discretion
Deterministic multi-step workflows live in workflows/ as YAML. A step can be a skill, an agent or a tool; steps are ordered with depends_on; data flows through ${{ }} templates; and a step can override its prompt. The README says every run follows the same path, with no LLM discretion on execution order.
In its example, named code-review-flow and triggered by pull_request, a lint step runs the static-analysis skill, a review step runs the code-reviewer agent after lint with a prompt telling it to flag eval() and raw SQL, a test step runs the bash tool with npm test and coverage, and a report step uses the review-summary skill after the review and test steps.
The trade is predictability against flexibility. A fixed graph is easy to audit and reproduce, but it cannot adapt its steps to what an earlier step discovers. The README's own list of items that stay in the framework includes iterative loops, so a workflow that needs to loop until a condition is met needs framework support beyond this file.
Git itself supplies versioning, review gates and secrets hygiene
Many of the listed patterns are git practices applied to agents. Every change to an agent is a commit, so you can roll back a broken prompt. git diff shows what changed between versions and git blame traces each line. Tagged releases such as v1.1.0 let you pin production to a tag and canary new versions on staging. Branches such as dev, staging and main can promote changes through environments.
For agents that learn, the human-in-the-loop pattern has them open a branch and a pull request before new skills or memory writes merge. Live memory sits in memory/runtime with dailylog.md, key-decisions.md and context.md. Secrets stay out of the repo: tools read API keys from a local .env that .gitignore keeps out of version control.
The limit is that all of this assumes the people and systems around the agent use git properly. A monorepo shares root-level context.md, skills/ and tools/ across every agent in it, which removes duplication but also means one edit affects all of them.
Pre-1.0 releases, a TypeScript build step and a small dependency list
The package is at version 0.5.0 and the last push was on 2026-07-02. The repository is not archived. Because the version is below 1.0, treat the CLI behaviour and the spec as able to change between minor versions, which the rename history already demonstrates.
The code is TypeScript. The scripts are build, which runs tsc and marks dist/index.js executable, dev for tsc --watch, start for node dist/index.js, and test, which runs node --test over compiled test files. Runtime dependencies are ajv and ajv-formats for schema validation, chalk, commander, inquirer and js-yaml. The published files are dist/, spec/, examples/, README.md and docs.md.
The repository holds a spec/ directory with SPECIFICATION.md, a patterns/ directory, and example agents including minimal, standard, full, financial-agent, llm-wiki, lyzr-agent and gitagent-helper.
Editorial conclusion
Use OpenGAP if you want agent prompts, rules and roles kept in git with diffs, branches and a validator, and you accept that the framework still runs the agent. Skip it if you expected a runtime or a finished adapter for your framework, since the README says runtime orchestration stays where it is. Before adopting it, remember it is at v0.5.0 and has been renamed twice, so check the package name and the spec version you are targeting.
Frequently asked questions
What does OpenGAP do?
OpenGAP, the Git Agent Protocol, is a framework-agnostic, git-native standard for defining AI agents. You clone a repo and get an agent definition, and opengap is its reference CLI.
How do I install OpenGAP?
The CLI is published on npm as @open-gitagent/opengap and requires Node 18 or later. It installs two commands, opengap and gitagent, the second kept as a backward-compatible alias.
Which files are required in an OpenGAP agent repository?
Only agent.yaml, the manifest, and SOUL.md, the identity. Files like RULES.md, DUTIES.md and AGENTS.md are optional.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/open-gitagent-opengap)Star history chart for your README
[](https://hysenlabs.com/projects/open-gitagent-opengap)Was this useful?