geml

GEML — General Expressive Markup Language

English | 中文

GEML is an Agent-Native base document format and protocol, designed for people and AI agents to read and write the same document.
One format, two readers. In agent-driven development and knowledge work, plain text and Markdown have no deterministic block boundaries: a program and a model trade the whole file in and the whole file back out — at best probing for it with line windows, and restating the original verbatim to rewrite it. Token cost grows with the length of the document, and the operation turns bloated. After a few rounds of rewriting, the copies excerpted elsewhere start to drift.

GEML organizes a document into typed blocks, each with a type and a unique id, so a model locates by #id and edits in place. With built-in write validation and .gemlhistory tracking, an agent reads and writes at very low token cost — leaving the precious context window for the actual work.

For people, it is plain text that reads clean; for agents, it is an addressable, verifiable, traceable, revertible "Doc-as-a-Base".

---

GEML is minimal. It is plain text — still clean with no renderer in sight; one block syntax for the whole language; addressable, verifiable, referenceable structure, natively.

Instead of a separate mini-syntax for each kind of content, GEML carries every kind in one container: the typed block. Code is a block. So are tables, diagrams, math, callouts, even metadata — and a run of prose can be one too (=== text), whenever you want it addressable. Extending it later is just as plain. The shape is the same every time, which makes the language easy enough to learn that it's hard to get wrong.

=== code {#hello lang=python}
print("hi")
===

geml get doc.geml '#hello'   # by name, just this block

Blocks have names so the verbs have somewhere to land — the full syntax is in the format in 1 minute.

Contents: What it solves · Why now · What's different · The format in 1 minute · A gift for programmers · Get hands-on · With an LLM · Maturity & versions · The design · Roadmap · Take part · License

What it solves

Why the LLM era needs a brand-new text format

Because both the producer and the consumer of a document have changed.

In traditional software engineering, a document was either a static explanation for people to read, or a serialized data file for programs.

Today, people and AI agents collaborate on the same document at high frequency. When the agent becomes the document's "second reader and co-author", the old balance breaks for good:

1. Context is scarce compute: every whole-document read or write burns an agent's limited attention window and reasoning budget;
2. Human–machine collaboration needs an isomorphic carrier: people need to read it at a glance, agents need to read and write it precisely, block by block;
3. Knowledge must have a single source of truth: scattered prompts and copy-pasted Markdown are destined to decay with every iteration.

Yet none of our existing text infrastructure was designed for this scene:

• Markdown (typeset for people): no stable structural blocks, no machine keys. To change one parameter, an agent must read and write the whole text — wasting context budget across multi-turn loops, and inviting drift in both format and meaning.
• JSON / XML (serialized for machines): full of wrapper syntax and structural noise — blocking natural human reading, while quietly eating expensive tokens in long contexts.
• Scratch memory and scattered files (no single source of truth): context is torn across chat history and Markdown copies everywhere; a copy is drift from the moment it is made, and version skew and hallucinated distortion follow.

The answer: "Doc-as-a-Base"

GEML invents no heavy new runtime. It gives plain-text documents one standard set of operational semantics:

A document no longer needs just a format — it needs a set of verbs. GEML keeps plain-text readability and adds deterministic block-level operations.

💡 Deep Dive: If you are interested in the dilemma of engineering documents in the LLM era and why we need to redesign a plain-text format from the ground up, read our full article on the blog: "Why Do We Need a New Text Format in the Era of LLMs?"

What's different about GEML

GEML stays small on purpose — the thinking, what it refuses, and what is still open are in how we thought about the design.

The four capabilities were established a chapter ago — addressing, projection, validation, rollback. This chapter is where each format lands against them, and where GEML draws its boundaries.

How other formats compare

Each of the four has mature solutions in its own field; what's unusual is meeting all four in one plain-text format:

Item by item: vs. CommonMark · vs. XML and JSON · a 7-format capability matrix.

Coexisting with Markdown: GEML is the editing source of truth, Markdown is the delivered artifact. Project one way with geml <file> --to md|html and ship .md or .html as before. Collaboration, not lock-in. (Projection is lossy: block ids and table-bound charts don't survive it.)

Don't take the table's word for it — re-run it. This is what I asked the model:

Based on your own experience editing the READMEs just now, describe the command steps you go through on a document (I saw you using grep and such), and whether you cache documents to save tokens — let's compare, and from that see which parts of GEML would actually earn their place.

What came back: what one edit costs and a real day replayed. Paste the question to your own model and see what it tells you. PS: I am still trying to work out whether the upstream chain (who calls this) and the downstream chain (what it calls) that codemap produces can pin down functions and call sites — and change project code — the same way. I will post a report when I have one.

The format in 1 minute

Typed blocks

One shape, every type. A block's basic syntax is === type [attributes] … === (where attributes like {#id .class key=val} are optional) — only the type (and how its body is read) changes:

=== code {lang=python}
print("hi")
===

=== note {.intro}
Parsed prose with *emphasis* and a [[#budget]] reference.
===

=== meta
title = "Budget plan"
===

A run of = (three or more) opens a block; an equal-length run closes it; longer fences nest inside shorter ones. A block that carries an #id can also close with the labeled fence === #id — no fence-length counting, which makes long blocks much harder to get wrong (nesting still requires a longer outer fence: a same-length bare === in the body closes the block early, labeled or not). The type decides how the body is read — raw (verbatim: code, diagram, math, table), flow (parsed prose with inline markup: note, text), or data (one key=val per line: meta); embed carries no body at all — its src= names the block it stands for — and every block may carry an attribute object {#id .class key=val}, where a .class is a semantic label, never a styling hook. The full inline grammar (emphasis, links, [[#id]] auto-references, media, footnotes, inline $math$) is in the spec.

Tables — two bodies, one model

Write a table visually:

=== table {#budget caption="Annual cost"}
| Plan  | Months | Rate |
|-------|-------:|-----:|
| Basic |      1 |   30 |
| Pro   |      2 |   30 |
===

…or as data, with computed columns and a summary row:

=== table {#fy25 format=csv header=1 compute="FY [%.1f] = Q1 + Q2 + Q3 + Q4" summary="Segment = 'Total'; FY [%.1f] = sum(FY)"}
Segment,  Q1, Q2, Q3, Q4
Cloud,     8, 10, 12, 14
Platform,  5,  6,  7,  9
Services,  3,  4,  4,  5
===

Both forms describe the same model. The FY column and Total row are computed at build time:

compute runs + - * / ( ) per row over columns; summary adds a foot row from the aggregates sum / avg / min / max / count (with arithmetic over them, e.g. weighted ratios); a trailing [printf] sets numeric display.

Tables can also pull their data from an external CSV via src="regions.csv".

❓ Up for discussion: should computed columns and the summary row stay? Keep, freeze, or drop — say which.

Math

=== math {#gauss caption="Gaussian integral"}
\int_{-\infty}^{\infty} e^{-x^2} dx = \sqrt{\pi}
===

$$\int_{-\infty}^{\infty} e^{-x^2} dx = \sqrt{\pi}$$

Diagrams & charts — host a DSL, or chart a table

GEML never interprets a diagram body; it routes it to a pluggable renderer (an unknown format is a warning, body preserved):

=== diagram {#flow format=mermaid caption="Review flow"}
graph LR
  A[Draft] --> B{Review} -->|ok| C[Publish]
===

graph LR
  A[Draft] --> B{Review} -->|ok| C[Publish]

A diagram can also chart a table — single source of truth, with the column references checked at build time and no data copied:

=== diagram {format=geml-chart data=#fy25 type=bar x=Segment y=FY}
===

Drawn from the #fy25 table above:

xychart-beta
  title "FY by segment"
  x-axis [Cloud, Platform, Services]
  y-axis "FY"
  bar [44, 27, 16]

Data — a value, not just text

Every block type names what it holds: code a region of code, table a grid, math a formula. data holds a data value, and it is where the data formats live — json (the default) and jsonl today, yaml/toml reserved. Being typed means the body is read, not just displayed: a missing comma fails the build, geml get --json returns the value itself, and a chart can read it directly.

=== data {#log format=jsonl}
{"ts":"09:00","p95":41}
{"ts":"09:10","p95":58}
===

A jsonl body holds one record per line, which a program can blind-append at end-of-file. Records can also stay in their own file: src=ops/latency.jsonl#L900-999 names the file and, optionally, a line window — so the log keeps being appended and tailed as before, while the document is its verified, addressable, chartable view of it.

Embeds — a dynamic reference, not a copy

One block can stand for another: in the same document by src=#id, across documents by src=other.geml#id. An embed is a dynamic lookup of the source at render time — change the source once and every embed follows; delete it and geml check fails the build on the spot.

=== embed {src=#fy25}
===

The body stays empty; the target lives in src=.

Markdown can't show you the projection. To see it live: install the browser extension, open the raw link to sample.geml, and scroll to the Transclusion section — a same-document projection (src=#roadmap), cross-document projections, and even chained resolution (an embed pulls a chart, which itself binds to a table in another file) all render in place: nothing is written there, yet edit the source once and the projection follows.

A gift for programmers — geml-code-graph

To really feel how powerful and flexible a single GEML primitive is, let's try it on a code graph — a familiar but demanding case for programmers: your whole codebase's call graph, written as GEML. geml codemap build lays the call graph out as a tree of GEML documents — every method an #id block, with #calls / #called-by edges both ways. The downstream chain (what a method calls) for troubleshooting, the upstream chain (who calls it) for the blast radius — all visible in a second;

npm i -g @geml/geml
geml codemap build              # --root defaults to . : detect languages -> index -> one merged graph in ./.geml-code-graph/
geml codemap serve              # opens your browser on the graph

[!NOTE] Requirements. Node 22+ for the CLI (npm i -g @geml/geml). Everything below is optional and used only where noted: Joern for non-TS/JS languages in the code graph, and Chrome for the viewer extension.

[!TIP] TS/JS — zero setup: build fetches the scip indexer by itself. Java / C / Python / Go / Kotlin — one extra download, Joern: unzip its release package and pass that folder to build, e.g. --joern ~/joern/joern-cli (--joern C:\joern\joern-cli on Windows), or put it on PATH and skip the flag. Mixed front-end + back-end repo — everything merges into one graph.

geml-code-graph is itself a diagram format — one line embeds it in any GEML document (=== diagram {format=geml-code-graph src=.geml-code-graph/index.geml} ===), and an optional per-commit hook (bundled with the Claude skill) rebuilds it as the code moves, so the graph doesn't drift.

Scale is measured, not promised: on Apache Flink's codebase — 13,585 Java source files, ~81,000 methods, 266,821 call edges — the plain-text data tables still open and query instantly, and you can grep any method name to trace its call chain. Reproduce it yourself: clone apache/flink and run geml codemap build --joern … at its root.

Next — get hands-on now

▶ Try writing GEML in the Playground — edit on the left, rendered live on the right, and the build verdict flips red the moment a reference breaks. No install, nothing to read first.

Then, in the order that suits you:

1. See it render in your browser. Install the extension and open a raw .geml link (the raw file, not the GitHub blob page — that one is HTML): the GEML spec itself (dogfood — the spec is a GEML document, rendered at scale), the showcase (a computed table, four charts, a Mermaid flow, and math), or playground/sample.geml for the interactive code-graph.
2. Run it locally. npm i -g @geml/geml (Node 22+), then geml check a document, or point it at your own repo with geml codemap build.
3. Set up Claude Code — one command. npx -y @geml/geml skill install puts the authoring skill, the CLI and the MCP server in place, user-global, for every project. It edits no settings and installs no hooks. Details.
4. Read the grammar. The full spec (EN / 中文) is normative and short enough to read in a sitting.

Using GEML with an LLM

The goal is one thing: your model edits a block at a time, and verifies — never re-reads and re-emits a whole file to change one paragraph. Getting there takes one step, and which step depends on what you use.

Using Claude Code — run this

npx -y @geml/geml skill install

It installs the authoring skill, the geml CLI and the MCP server, user-global, for every project. No settings.json edits, no hooks; re-run after an upgrade. (Prefer plugins? claude plugin marketplace add geml-spec/geml, then /plugin install geml@geml — same skill, MCP server bundled.)

Using DeepSeek Harness — add this bundle

The same setup, packaged as a dsh bundle — the geml MCP server plus the authoring and code-graph skills:

dsh plugin --profile web add @geml/dsh-plugin   # web = the profile dsh boots by default; use your own profile name if you run another

Listed on dshmarket and awesome-dsh-plugin; source in integrations/dsh-plugin/.

Then say it once in a session, and the project has switched:

This project uses GEML as its base document format; generate other formats from it as needed.

The skill takes it from there. New documents are written as .geml with an id on every section — that id is what later lets one section be replaced instead of the file. Documents that already exist are left where they are: adopting the format is not licence to convert or delete anything. And geml <file> --to md|html produces whatever still has to ship as something else.

Using anything else — paste this, then check the output

A model with no skill to read needs the rules once. Paste the prompt below, and keep geml check as the gate on whatever it writes back — the CLI is npm i -g @geml/geml (Node 22+).

Write the document as GEML: every block is === type [attributes] … === (the format in 1 minute lists the types). Four rules are the ones models get wrong: the closing fence is a = run of the exact opening length, and a body containing === needs a longer outer fence; headings are ATX # only, with no --- frontmatter (metadata is === meta); every #id is unique and every reference ([[#id]], [text](#id), [^id], data=#id) must resolve; there is no raw HTML. The normative spec is GEML-spec.md.

What it will do with it

geml list   doc.geml                                     # CALL FIRST: every block, its address, kind, lines
geml find   "words" doc.geml                             # search block content -> an address, not a line number
geml get    doc.geml '#hello'                            # read ONE block (a heading id = its whole section)
geml get    doc.geml '#hello' --intro                    # a section cuts three ways: --head | --intro | --body
geml set    doc.geml '#license' --in template.geml#mit   # replace that block, forking another
geml add    doc.geml --after '#intro' --in snippet.geml  # insert a fragment (keeps its own ids)
geml revert doc.geml '#plan' --rev -1                    # roll ONE block back
geml check  doc.geml                                     # validate only: diagnostics + exit code

Any section cuts three ways, on get and set alike: --head is the heading line, --intro what it says before its first subheading, --body everything under it — so --body always contains --intro, and equals it when there is no subheading. A section's opening can be edited without pulling its subsections into context.

Every mutation is re-parsed before it writes and refused if it would break the document — which is what makes editing unattended safe. The rest of the verbs (delete, rename, history, --to md|html|geml conversion, addressing a block by type or content hash) are in the parser README.

MCP Server

A standard Model Context Protocol server ships with the package, so your agent edits one block at a time instead of rewriting whole files. It runs locally on Windows, macOS, and Linux; --root is the directory holding your .geml files.

Claude Code / any CLI client — one command:

claude mcp add geml -- npx -y @geml/geml@latest mcp --root /absolute/path/to/your/docs

Claude Desktop — add to claude_desktop_config.json:

View on GitHub