Skip to content

AICMO/llm-call-action

Use this GitHub action with your project
Add this Action to an existing workflow or create a new one
View on Marketplace
BranchesTags

Repository files navigation

LLM Call Action

prompt + data → LLM → parsed output

A reusable composite GitHub Action that takes a prompt template and data file, calls an LLM, and returns cleaned output. Single bundled TypeScript runtime — no external dependencies at execution time.

Quick Start

The fastest way to get started — use your personal Claude Code OAuth token:

- uses: aicmo/llm-call-action@v1
  with:
    prompt_file: prompts/summarize.md
    data_file: /tmp/input-data.txt
    claude_code_oauth_token: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}

- run: cat /tmp/llm_response.txt

No API keys, no billing setup, no provider configuration. Just your personal token and a prompt. If the token is unavailable or you need a different provider, the action automatically falls back to a configured API key — just add one as an optional safety net.

How to get your Claude Code OAuth token

  1. Install Claude Code CLI
  2. Run claude setup-token — it will open a browser, authenticate, and print a long-lived token
  3. Copy the token and add it as a repository secret: Settings → Secrets → New repository secret → name it CLAUDE_CODE_OAUTH_TOKEN

Usage policy: Personal Claude Code tokens are tied to your individual Anthropic account. Anthropic's Terms of Service permit personal use. Business or commercial use may require a separate arrangement — it is your responsibility to ensure compliance with Anthropic's current terms. This project does not grant any license to use Anthropic's services.

Why this action?

Most CI/AI actions are either vendor-locked agents that own the whole workflow (PR comments, commits, code review) for a single provider, or thin API wrappers that hand you raw JSON and leave the rest to you.

This action is neither. It's a composable pipeline primitive — the curl of LLM workflows:

  • You control the prompt (inline text or template file).
  • You control the data (any file piped through data_file).
  • You decide what to do with the output (write to file, post a comment, feed into the next step).

The action handles the undifferentiated work: provider abstraction, auth routing, code-fence stripping, error detection, and response validation. You chain the steps.

Vendor agents Generic wrappers LLM Call Action
Multi-provider No (1 vendor) Yes Yes (4 providers)
OAuth + API unified No No Yes (auto-routed)
Pipeline-oriented No No Yes (output_file → next step's data_file)
Output cleaning No (raw response) No (raw response) Yes (code-fence stripping, error detection)
Zero runtime overhead No (installs deps) Varies Yes (single bundled file)
Opinionated about workflow Yes (PR comments, commits) No No — you decide

Use vendor agents when you want an AI that reads your repo, writes code, and posts PR comments. Use this action when you want a reliable, provider-agnostic building block that fits into any workflow you design.

Authentication

Personal Claude Code token (recommended)

Pass your personal OAuth token. Under the hood, this delegates to anthropics/claude-code-action@v1 — same engine that powers Claude Code CLI.

- uses: aicmo/llm-call-action@v1
  with:
    prompt_file: prompts/summarize.md
    data_file: /tmp/input-data.txt
    claude_code_oauth_token: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}

This is the preferred path: zero API billing, leverages your existing Claude Pro/Team subscription, and supports multi-turn tool use via max_turns.

API key (fallback)

When no OAuth token is provided, the action falls back to a direct API call via fetch. Supports 4 providers — pass the corresponding key:

# Claude (Anthropic API)
- uses: aicmo/llm-call-action@v1
  with:
    prompt: "Summarize the following data."
    data_file: /tmp/input-data.txt
    provider: claude
    anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
# OpenAI
- uses: aicmo/llm-call-action@v1
  with:
    prompt: "Summarize the following data."
    data_file: /tmp/input-data.txt
    provider: openai
    openai_api_key: ${{ secrets.OPENAI_API_KEY }}
# Gemini
- uses: aicmo/llm-call-action@v1
  with:
    prompt: "Summarize the following data."
    data_file: /tmp/input-data.txt
    provider: gemini
    google_api_key: ${{ secrets.GOOGLE_API_KEY }}
# Vertex AI
- uses: aicmo/llm-call-action@v1
  with:
    prompt: "Summarize the following data."
    data_file: /tmp/input-data.txt
    provider: vertex
    google_application_credentials_json: ${{ secrets.GOOGLE_CREDENTIALS_B64 }}
    vertex_project: my-gcp-project

Features

  • 2 auth paths — Personal Claude Code token (recommended) or API key per provider
  • 4 API providers — Claude, OpenAI, Gemini, Vertex AI
  • Inline or file prompts — pass prompt text directly or reference a template file
  • Response cleaning — automatic output sanitization (see below)
  • Pipeline-ready — output_file of one step becomes data_file of the next
  • Prompt stats — logs character/word/line counts for debugging

Response Cleaning

Enabled by default (clean_response: true). Set to false for raw passthrough.

When enabled, the action applies three cleaning steps to every LLM response before writing it to output_file:

  1. Code-fence stripping — removes opening (```lang) and closing (```) fence lines, so you get plain text instead of a markdown code block.
  2. Error detection — scans the response for known failure patterns and fails the step if any match:
    • I cannot, I'm unable to
    • ENOENT, No such file
    • error occurred
    • cannot be accessed
    • outside the allowed working directory
  3. Minimum length validation — fails if the cleaned response is shorter than min_response_length characters (default: 50). Catches empty or stub responses.

If any check fails, the action exits with a non-zero code and logs the first 5 lines of the response as a preview.

Inputs

Input Required Default Description
prompt No* — Inline prompt text
prompt_file No* — Path to prompt template file
data No — Inline data text to append to prompt
data_file No — Path to data file to append to prompt
output_file No /tmp/llm_response.txt Where to write the parsed LLM response
claude_code_oauth_token No — Personal Claude Code OAuth token (recommended — enables OAuth path)
max_turns No 5 Max turns for Claude Code OAuth path
claude_tools No — Restrict available tools for Claude Code OAuth path (e.g., Read, Bash,Edit,Read). Leave empty for all tools.
provider No claude LLM provider for API path: claude, openai, gemini, vertex
model No Per-provider default Model name (provider-specific)
max_tokens No 4096 Max tokens for LLM response
clean_response No true Clean the response: strip code fences, detect error patterns, validate min length
min_response_length No 50 Minimum response length in characters (only when clean_response is true)
anthropic_api_key No — Anthropic API key
openai_api_key No — OpenAI API key
google_api_key No — Google AI API key (Gemini)
google_application_credentials_json No — Google service account credentials (base64-encoded, Vertex AI)
vertex_project No — Google Cloud project for Vertex AI
vertex_region No us-central1 Google Cloud region for Vertex AI

* Either prompt or prompt_file must be provided.

Default Models

Provider Default Model
claude (OAuth) claude-sonnet-4-6
claude (API) claude-sonnet-4-6
openai gpt-4o
gemini gemini-2.5-pro
vertex gemini-2.5-pro

Outputs

Output Description
response The parsed text response from the LLM
response_file Path to the file containing the parsed response

Scenarios

Content Generation

Release Notes

- run: git log --oneline v1.0.0..HEAD > /tmp/git-log.txt

- uses: aicmo/llm-call-action@v1
  with:
    prompt_file: prompts/release-notes.md
    data_file: /tmp/git-log.txt
    claude_code_oauth_token: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}

Documentation Regen

- run: cat src/api/*.ts > /tmp/api-source.txt

- uses: aicmo/llm-call-action@v1
  with:
    prompt_file: prompts/generate-docs.md
    data_file: /tmp/api-source.txt
    claude_code_oauth_token: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}

Code Quality

PR Review

- run: gh pr diff ${{ github.event.pull_request.number }} > /tmp/pr-diff.txt

- uses: aicmo/llm-call-action@v1
  with:
    prompt_file: prompts/pr-review.md
    data_file: /tmp/pr-diff.txt
    claude_code_oauth_token: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}

Data Transformation

i18n Translation

- uses: aicmo/llm-call-action@v1
  with:
    prompt_file: prompts/translate-to-es.md
    data_file: locales/en.json
    output_file: locales/es.json
    claude_code_oauth_token: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}

Multi-step Pipelines

Analyze → Summarize → Recommend

- uses: aicmo/llm-call-action@v1
  with:
    prompt_file: prompts/analyze.md
    data_file: /tmp/raw-data.txt
    output_file: /tmp/analysis.txt
    claude_code_oauth_token: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}

- uses: aicmo/llm-call-action@v1
  with:
    prompt_file: prompts/summarize.md
    data_file: /tmp/analysis.txt
    output_file: /tmp/summary.txt
    claude_code_oauth_token: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}

- uses: aicmo/llm-call-action@v1
  with:
    prompt_file: prompts/recommend.md
    data_file: /tmp/summary.txt
    output_file: /tmp/recommendations.txt
    claude_code_oauth_token: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}

Mixed Providers

Nothing stops you from mixing auth paths or providers within the same workflow:

# Step 1: Claude via personal token
- uses: aicmo/llm-call-action@v1
  with:
    prompt_file: prompts/analyze.md
    data_file: /tmp/raw-data.txt
    output_file: /tmp/analysis.txt
    claude_code_oauth_token: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}

# Step 2: OpenAI via API key
- uses: aicmo/llm-call-action@v1
  with:
    prompt_file: prompts/summarize.md
    data_file: /tmp/analysis.txt
    output_file: /tmp/summary.txt
    provider: openai
    openai_api_key: ${{ secrets.OPENAI_API_KEY }}

Architecture

Execution Flow

                    ┌─────────────────────┐
                    │  Step 1: Build      │
                    │  prompt-builder.js   │
                    └──────────┬──────────┘
                               │ writes .llm_user_prompt.txt
                               │
                    ┌──────────▼──────────┐
                    │  Step 2: Auth       │
                    │  check (bash)       │
                    └───┬─────────────┬───┘
                        │             │
                 oauth  │             │  api
                        │             │
           ┌────────────▼───┐   ┌─────▼──────────────┐
           │ Step 3a:       │   │ Step 3b:            │
           │ claude-code-   │   │ api-caller.js       │
           │ action@v1      │   │ (fetch → provider)  │
           └──────┬─────────┘   └─────┬──────────────┘
                  │ exec_file         │ /tmp/llm_raw.txt
                  │                   │
                  └─────────┬─────────┘
                            │
                   ┌────────▼─────────┐
                   │ Step 4: Parse    │
                   │ response-parser  │
                   └────────┬─────────┘
                            │
                     sets outputs:
                     • response (text)
                     • response_file (path)

Steps 3a and 3b are mutually exclusive — the auth check routes to exactly one. Step 4 auto-detects which path ran by checking which file exists on disk.

Step-by-Step Detail

Step 1 — Build Prompt (dist/prompt-builder.js)

Runs: Always.

Resolves prompt content from one of two sources, optionally appends data, writes the assembled prompt to $GITHUB_WORKSPACE/.llm_user_prompt.txt.

Env var Source
INPUT_PROMPT inputs.prompt
INPUT_PROMPT_FILE inputs.prompt_file
INPUT_DATA_FILE inputs.data_file
GITHUB_WORKSPACE Set by GitHub Actions

Logic:

  1. If INPUT_PROMPT is non-empty → use as prompt content.
  2. Else if INPUT_PROMPT_FILE is non-empty → read that file.
  3. Else → fail.
  4. If INPUT_DATA_FILE is set → read and append with \n---\nData:\n delimiter.
  5. Write to .llm_user_prompt.txt. Log stats (chars, words, lines, head/tail preview).

Produces: $GITHUB_WORKSPACE/.llm_user_prompt.txt

Failure cases:

  • Neither prompt nor prompt_file provided
  • Prompt file or data file does not exist

Step 2 — Auth Check (bash)

Runs: Always.

One-liner: if claude_code_oauth_token is non-empty, output method=oauth; otherwise method=api.

Produces: steps.auth.outputs.method ("oauth" or "api")

Step 3a — OAuth Path (anthropics/claude-code-action@v1)

Runs: Only when method == 'oauth'.

Delegates to the official Claude Code Action using your personal token. Passes a prompt that tells Claude to read .llm_user_prompt.txt and respond with raw output only.

Parameter Value
claude_code_oauth_token From input
prompt "Read .llm_user_prompt.txt, respond with ONLY the output"
--model inputs.model or claude-sonnet-4-6
--max-turns inputs.max_turns or 5
--tools inputs.claude_tools (omitted if empty — all tools available)

Produces: steps.llm_oauth.outputs.execution_file — a JSON file containing the execution log (array of entries with type: "result").

Failure: Inherits from claude-code-action (auth failure, model error, etc.).

Step 3b — API Path (dist/api-caller.js)

Runs: Only when method == 'api'.

Reads the assembled prompt, resolves the provider, validates auth, and calls the LLM API via native fetch.

Env var Source
INPUT_PROVIDER inputs.provider (default: claude)
INPUT_MODEL inputs.model (default: per-provider, see table above)
INPUT_MAX_TOKENS inputs.max_tokens (default: 4096)
INPUT_ANTHROPIC_API_KEY inputs.anthropic_api_key
INPUT_OPENAI_API_KEY inputs.openai_api_key
INPUT_GOOGLE_API_KEY inputs.google_api_key
INPUT_GOOGLE_APPLICATION_CREDENTIALS_JSON inputs.google_application_credentials_json
INPUT_VERTEX_PROJECT inputs.vertex_project
INPUT_VERTEX_REGION inputs.vertex_region (default: us-central1)

Provider details:

Provider Endpoint Auth Response extraction
claude api.anthropic.com/v1/messages x-api-key header .content[0].text
openai api.openai.com/v1/chat/completions Bearer token .choices[0].message.content
gemini generativelanguage.googleapis.com/.../generateContent API key as ?key= query param .candidates[0].content.parts[0].text
vertex {region}-aiplatform.googleapis.com/.../generateContent JWT → OAuth2 token exchange → Bearer .candidates[0].content.parts[0].text

Produces: /tmp/llm_raw.txt (extracted text content from the API response).

Failure cases:

  • Unknown provider name
  • Missing API key / credentials for the selected provider
  • API returns an error response (auth failure, rate limit, etc.)
  • API returns a response with no extractable content
  • Vertex-specific: invalid base64 credentials, missing project, token exchange failure

Step 4 — Parse Response (dist/response-parser.js)

Runs: Always (after whichever Step 3 ran).

Reads the raw LLM output, cleans it, validates it, and sets the action's outputs.

Env var Source
INPUT_OUTPUT_FILE inputs.output_file (default: /tmp/llm_response.txt)
INPUT_CLEAN_RESPONSE inputs.clean_response (default: true)
INPUT_MIN_RESPONSE_LENGTH inputs.min_response_length (default: 50)
LLM_RESPONSE_CLAUDE_EXEC_FILE steps.llm_oauth.outputs.execution_file (empty if API path)

Logic:

  1. Find input (priority order):
    • If LLM_RESPONSE_CLAUDE_EXEC_FILE exists → read it, parse as JSON array, extract the entry where type === "result" → use .result.
    • Else if /tmp/llm_raw.txt exists → read it directly.
    • Else → fail: no input source.
  2. If empty or whitespace-only → fail.
  3. If clean_response is true (default):
    • Strip code fences — remove lines matching ```language or ```.
    • Error detection — regex scan for patterns: I cannot, I'm unable to, ENOENT, No such file, error occurred, cannot be accessed, outside the allowed working directory. If matched → fail with the first 5 lines as preview.
    • Length validation — if content.trim().length < min_response_length → fail.
  4. If clean_response is false — write raw content as-is (skip steps above).
  5. Write content to INPUT_OUTPUT_FILE.
  6. Set outputs: response (text) and response_file (path to the written file).

Produces:

  • File at output_file path
  • Action output response (text)
  • Action output response_file (path)

Failure cases (always):

  • No input source (neither Claude exec file nor /tmp/llm_raw.txt exists)
  • Empty content from LLM

Failure cases (only when clean_response: true):

  • Error pattern detected in the response
  • Response shorter than min_response_length characters

Development

npm install
npm run build      # Rollup → 3 dist/*.js files
npm test           # Vitest unit tests
npm run typecheck  # TypeScript type checking

About

Github Action to call LLMs. With Logic 1 - build prompt (with data), 2 - call LLM (trying personal first, fallback to API), 3 - parse response (free text or JSON)

Resources

Readme

License

Apache-2.0 license
Activity
Custom properties

Stars

1 star

Watchers

0 watching

Forks

0 forks
Report repository

Releases 6

v1.1.4 Latest
Apr 3, 2026
+ 5 releases

Packages

 
 
 

Contributors

Languages