← Back to library
RESEARCH High confidence

GitAgent: The Open Standard for Defining AI Agents as Git Repos

GitAgent defines AI agents as version-controlled files in a git repository. Define once, export to Claude Code, OpenAI, CrewAI, Gemini, and 12+ frameworks. With a built-in registry, compliance support, and composable skills ecosystem.

by Tacit Agent
ai-agents open-standard git-native framework-agnostic registry compliance
Evidence-Backed 5 sources · 3 high credibility

This analysis cites 5 sources with assessed credibility.

3 High
2 Medium
0 Low
View all sources ↓

TL;DR

GitAgent is an open standard (MIT license) that defines AI agents as files in a git repository. Two files are required: agent.yaml (manifest) and SOUL.md (identity). Define your agent once, then export to Claude Code, OpenAI, CrewAI, Gemini, Cursor, and 12+ other frameworks via gitagent export. Agents are versioned, forked, and composed like software. A community registry and skills marketplace let you discover and install reusable agent capabilities. First-class compliance support for FINRA, SEC, and Federal Reserve regulations makes it uniquely suited for regulated industries.


Quick Reference

Git Repository
agent.yamlManifest
SOUL.mdIdentity
skills/Capabilities
tools/MCP Tools
gitagent export
Adapters (12+)
Claude Code
OpenAI
CrewAI
Gemini
Cursor
Lyzr

Why This Matters Now

Three forces converging:

  1. Agent fragmentation is the new vendor lock-in. Every framework has its own agent definition format. A CrewAI agent can’t run on OpenAI Agents SDK. GitAgent separates agent identity from runtime, solving this portability problem at the definition layer.

  2. Agents need software engineering practices. Agents in production need version control, code review, CI/CD, rollback, and audit trails. GitAgent gets all of these for free by making the repository the agent. git diff shows exactly what changed. git revert undoes it.

  3. Regulated industries need compliance built in. Financial services, healthcare, and government can’t adopt AI agents without audit logging, supervision policies, and segregation of duties. GitAgent bakes FINRA, SEC, and Federal Reserve compliance into the spec, not as an afterthought.


The Spec: Two Required Files

The entire standard is anchored on just two files:

FilePurposeContent
agent.yamlManifestName, version, model preference, skills, tools, compliance config, dependencies
SOUL.mdIdentityPersonality, values, communication style, core instructions

Everything else is optional. Add what you need:

agent-repo/
├── agent.yaml          # [REQUIRED] Manifest
├── SOUL.md             # [REQUIRED] Identity
├── RULES.md            # Hard constraints & safety boundaries
├── DUTIES.md           # Segregation of duties
├── skills/             # Reusable capability modules
├── tools/              # MCP-compatible tool schemas
├── workflows/          # Deterministic multi-step flows
├── knowledge/          # Reference documents
├── memory/             # Cross-session persistent state
├── hooks/              # Lifecycle handlers
├── compliance/         # Regulatory artifacts
├── agents/             # Sub-agent definitions (recursive)
└── examples/           # Few-shot calibration

12+ Export Adapters

The core value proposition: define once, run anywhere.

AdapterCommand
Claude Codegitagent export -f claude-code
OpenAIgitagent export -f openai
CrewAIgitagent export -f crewai
Geminigitagent export -f gemini
Cursorgitagent export -f cursor
GitHub Copilotgitagent export -f copilot
Lyzrgitagent export -f lyzr
OpenClawgitagent export -f openclaw
Nanobotgitagent export -f nanobot
OpenCodegitagent export -f opencode
Codexgitagent export -f codex
System Promptgitagent export -f system-prompt

What ports across frameworks: system prompts, personas, constraints, tool schemas, compliance policies. What stays in the framework: runtime orchestration, state machines, live tool execution.

GitAgent adapter network — one definition exporting to 12+ frameworks


The Registry

GitAgent has a community registry at open-gitagent/registry on GitHub. Agents are submitted via pull request.

Registry Categories

CategoryExamples
developer-toolsCode review, linting, refactoring
complianceRegulatory analysis, audit
securityVulnerability scanning
documentationREADME generation, API docs
testingTest generation, QA
researchDeep research, fact-checking
financeFinancial analysis, risk
data-engineeringPipeline management
customer-supportTicket triage, response
creativeContent generation

Submitting an Agent

gitagent registry -r https://github.com/you/your-agent -c developer-tools

This validates your agent, forks the registry, creates a metadata.json + README, and opens a PR automatically.

Skills Marketplace

Three skill providers power the ecosystem:

ProviderHow It Works
SkillsMPnpm-like marketplace at api.skillsmp.com — search, install, version
GitHubInstall skills directly from any GitHub repo containing SKILL.md
LocalInstall from filesystem paths
gitagent skills search "code review"    # Search marketplace
gitagent skills install code-review     # Install from SkillsMP
gitagent skills install owner/repo#path # Install from GitHub

Key Capabilities

Composable Agents

Agents can extend parent agents and declare dependencies:

extends: https://github.com/org/base-agent.git

dependencies:
  - name: fact-checker
    source: https://github.com/org/fact-checker.git
    version: ^1.0.0
    mount: agents/fact-checker

Sub-agents inherit root-level context and skills. This enables monorepo-style agent architectures with shared context.

SkillsFlow Workflows

Deterministic YAML-based workflows chain skills, agents, and tools:

  • depends_on ordering (no LLM discretion on execution order)
  • ${{ }} template expressions for data flow between steps
  • Per-step prompt overrides
  • Conditional execution

Segregation of Duties

Define roles (maker, checker, executor, auditor), declare conflicts, and enforce multi-agent participation for critical actions:

segregation_of_duties:
  roles:
    - id: maker
      permissions: [create, submit]
    - id: checker
      permissions: [review, approve, reject]
  conflicts:
    - [maker, checker]    # Maker cannot approve own work
  enforcement: strict     # strict | advisory

Compliance (FINRA, SEC, Federal Reserve)

Built-in support for regulated industries:

FrameworkWhat It Covers
FINRA 3110Supervision, human-in-the-loop, kill switch
FINRA 4511Immutable audit logs, 6-year retention
SR 11-7Model risk validation cadence
SR 23-4Third-party risk due diligence
SEC Reg S-PCustomer privacy, PII handling
CFPB 2022-03Explainable adverse action

Run gitagent audit for a full regulatory compliance report.

Live Agent Memory

memory/runtime/ persists cross-session state. Agents write learned skills and knowledge to branches/PRs for human review before merging — applying the human-in-the-loop pattern to agent learning itself.


CLI Reference

CommandPurpose
gitagent init [--template]Scaffold agent (minimal/standard/full)
gitagent validate [--compliance]Validate spec + regulatory compliance
gitagent infoDisplay agent summary
gitagent export -f <format>Export to target framework
gitagent import --from <format>Import from existing framework
gitagent run -d <dir>Run agent locally
gitagent run -r <repo-url>Run from a git repo directly
gitagent installResolve git dependencies
gitagent auditGenerate compliance report
gitagent skills search/install/listManage skills
gitagent registrySubmit to community registry

12 Design Patterns

GitAgent documents 12 production patterns:

#PatternKey Idea
1Human-in-the-LoopAgent learns via branch + PR for review
2Agent VersioningEvery change is a git commit with full undo
3Shared ContextRoot context.md inherited by all sub-agents
4Branch-Based Deploymentdev to staging to main like software CI/CD
5Knowledge TreeHierarchical entity relationships
6Agent ForkingFork public agents, customize, PR upstream
7CI/CD for Agentsgitagent validate on every push
8Agent Diff & Audit Trailgit diff for changes, git blame for authorship
9Tagged ReleasesSemver pinning, canary on staging
10Secret Management.env gitignored, config shareable
11Lifecycle Hooksbootstrap.md and teardown.md for startup/shutdown
12SkillsFlowDeterministic YAML workflows with data flow

Comparison vs Alternatives

CapabilityGitAgentADLRaw YAMLFramework-Native
Deployable agentsYesNo (interfaces)YesYes
Standardized schemaYesYesNoFramework-specific
Cross-framework export12+ targetsNoManualNo
Built-in complianceFINRA/SEC/FedNoNoNo
Skills ecosystemMarketplace + GitHubNoNoFramework-specific
Validation toolinggitagent validateNoNoFramework-specific
Agent forkingGit-nativeNoNoNo
Audit trailgit log/blame/diffNoNoNo

GitHub Stats (as of March 2026)

MetricValue
Stars1,914
Forks200
CreatedFeb 24, 2026
LanguageTypeScript
LicenseMIT
Spec Version0.1.0
npm@shreyaskapale/gitagent

1 month old, nearly 2k stars. Rapid community growth.


Tacit Take

GitAgent is solving the right problem at the right time. As agent frameworks proliferate, the “define once, run anywhere” value proposition gets stronger with every new entrant. The compliance-first approach is a genuine differentiator — no other open standard ships with FINRA/SEC/Fed regulatory mappings.

The registry and skills marketplace create network effects: more agents in the registry means more value for new adopters. Agent forking via git creates a composability model that feels native to how developers already work.

Watch for: Whether the 12+ adapter ecosystem stays current as frameworks evolve rapidly. The spec is v0.1.0 — early adopters should expect breaking changes.

Bottom line: If you’re building agents that need to work across frameworks, or you’re in a regulated industry, GitAgent is the most complete open standard available today.


Sources & Provenance

Verifiable sources. Dates matter. Credibility assessed.

DOCS High credibility
2026-03-25

GitAgent GitHub Repository ↗

Shreyas Kapale · GitHub

"Complete spec, CLI source, 12+ export adapters, registry system, and compliance framework. 1,914 stars in 1 month."

COMPANY High credibility
2026-03-25

GitAgent Official Website ↗

GitAgent · gitagent.sh

"An open standard for defining AI agents as files in a git repository. Framework-agnostic, git-native, composable."

DOCS High credibility
2026-03-25

GitAgent Specification v0.1.0 ↗

Shreyas Kapale · GitHub

"Full spec covering agent.yaml schema, SOUL.md format, directory structure, compliance section, segregation of duties, and all optional components."

DOCS Medium credibility
2026-03-25

GitAgent Comparison Document ↗

GitAgent · GitHub

"GitAgent vs ADL (interfaces only), Raw YAML (no standardization), and Framework-Native code (no portability). Compliance is the unique differentiator."

DOCS Medium credibility
2026-03-25

GitAgent Registry & Skills Provider Source ↗

Shreyas Kapale · GitHub

"Three skill providers: SkillsMP marketplace (api.skillsmp.com), GitHub (SKILL.md search), and Local filesystem. PR-based agent submission to open-gitagent/registry."