The grep-native data language for agentic systems

GDL is a data language designed for how agents actually work - using native bash tools like grep. Seven grep-native file formats give AI agents structured knowledge about codebases, schemas, APIs, architecture, and decisions. The filesystem is the coordination layer, git is the audit trail, and grep is the query engine.

No databases. No message queues. No vector databases required.

The Seven Layers

Layer	Extension	Purpose
GDL (Data)	.gdl	Structured business data as @type|key:value records
GDLS (Schema)	.gdls	Structural maps of external systems (tables, columns, PKs)
GDLC (Code)	.gdlc	Code structure maps (modules, functions, dependencies)
GDLA (API)	.gdla	API contract maps (endpoints, schemas, auth)
GDLD (Diagram)	.gdld	Visual knowledge - flows, patterns, sequences, gotchas
GDLM (Memory)	.gdlm	Shared agent knowledge with three-tier lifecycle
GDLU (Unstructured)	.gdlu	Document indexes for PDFs, transcripts, media

All formats share: @ prefix, | delimiter, one record per line — grep works across all of them.

Example

@customer|id:C-001|name:Acme Corp|tier:enterprise|region:APAC
@T orders|Order headers
  customer_id|BIGINT|N|FK|Owning customer
  order_date|DATE|N||Date placed
@EP GET /pets/{id}|desc:Get pet by ID|bearer
@memory|id:M-001|agent:arch-review|type:decision|subject:Chose event sourcing over CRUD
@node|id:AuthService|label:Auth Service|group:backend

# Find all enterprise customers
grep "tier:enterprise" data/*.gdl

# List every table across all schemas
grep "^@T " schema/**/*.gdls

# Search agent memory for architecture decisions
grep "type:decision" memory/active/*.gdlm

Performance

Benchmarking is in progress. Early results show:

• Significantly smaller file sizes than YAML and JSON for equivalent data
• Fewer tokens per query compared to structured alternatives
• High accuracy at scale for table/column navigation tasks

Full benchmark results, methodology, and release notes are published at greppable.ai/benchmarks.

Documentation

Document	Description
specs/	Format specifications for all 7 layers
ARCHITECTURE.md	Core architecture, concurrency model, agent coordination
PROMPTS.md	Optimized minimal agent prompts per layer
docs/reference/	Grep pattern quick-reference guides (all 7 layers)

Contributing

See CONTRIBUTING.md for guidelines on proposing spec changes.

License

MIT - see LICENSE.