Reusable analysis templates for humans and machines
Package each analysis step as a template with explicit inputs and outputs. Connect steps through bindings. Run everything from a short CLI, a local JSON API, or an MCP agent.
pipx install git+https://github.com/chaochungkuo/linkar.git $ linkar config author set --name "Your Name"
$ linkar project init --name study
$ cd study
$ linkar pack add ../examples/packs/basic --id basic
$ linkar run simple_echo --name Linkar
$ linkar inspect run simple_echo_001
study/
simple_echo/ ← stable project alias
.linkar/runs/
simple_echo_001/
meta.json ← provenance
results/greeting.txt
How It Works
Every Linkar project is built from templates, bindings, and a project directory. Each piece has a single job — together they make analysis work reusable and inspectable.
Standalone units with explicit params, declared outputs, and a reproducible run command. Each template is testable on its own.
id: fastq_stats
params:
input:
type: path
required: true
outputs:
report:
path: stats.txt
run:
command: fastqc "${param:input}"Encode how outputs from one template become inputs to the next. Write the resolution logic once; reuse it every run.
# step_a produced output_file
# binding auto-resolves it
$ linkar run step_b
# no manual path wiring needed
A local directory with stable run aliases, immutable history under
.linkar/runs/, and inspectable metadata for every execution.
$ linkar inspect run fastqc_001
instance: fastqc_001
status: completed
outputs:
report: ./fastqc/stats.txtInterfaces
Humans use the CLI. Scripts and agents use the local JSON API or MCP server. All three paths write the same run artifact on disk.
Template params become flags. The active project is discovered automatically. Common paths stay short.
linkar config author set --name "Your Name"
linkar project init --name study
cd study
linkar pack add ../examples/packs/chaining --id chaining
linkar run produce_message --message "hello"
linkar run consume_message
linkar inspect run consume_message_001Resolve params, execute templates, and read outputs without shell scraping. Agents can use the MCP server directly.
linkar serve --port 8000 --api-token local-dev:read,resolve,execute
# resolve params → get a short-lived token
POST /v1/templates/{template_id}:resolve
# run with the token
POST /v1/templates/{template_id}:run
# read outputs by run reference
GET /v1/runs/{run_ref}/outputs
# or use the MCP server for agents
linkar mcp serveDocumentation
Read the concepts first to understand the model. Use the tutorials to do real work step by step. Browse the bundled examples when you want the smallest runnable pack for a specific pattern.