Input Formats

intake supports 12 input formats through specialized parsers. The format is auto-detected by file extension and content. Parsers are automatically discovered via the plugin system.

Summary table

Format	Parser	Extensions / Source	Dependency	What it extracts
Markdown	`MarkdownParser`	`.md`, `.markdown`	—	YAML front matter, sections by headings
Plain text	`PlaintextParser`	`.txt`, stdin (`-`)	—	Paragraphs as sections
YAML / JSON	`YamlInputParser`	`.yaml`, `.yml`, `.json`	—	Top-level keys as sections
PDF	`PdfParser`	`.pdf`	pdfplumber	Text by page, tables as Markdown
DOCX	`DocxParser`	`.docx`	python-docx	Paragraphs, tables, metadata, sections by headings
Jira	`JiraParser`	`.json` (auto-detected)	—	Issues, comments, links, labels, priority
Confluence	`ConfluenceParser`	`.html`, `.htm` (auto-detected)	bs4, markdownify	Clean content as Markdown
Images	`ImageParser`	`.png`, `.jpg`, `.jpeg`, `.webp`, `.gif`	LLM vision	Visual content description
URLs	`UrlParser`	`http://`, `https://`	httpx, bs4, markdownify	Web page content as Markdown
Slack	`SlackParser`	`.json` (auto-detected)	—	Messages, threads, decisions, action items
GitHub Issues	`GithubIssuesParser`	`.json` (auto-detected)	—	Issues, labels, comments, cross-references
GitLab Issues	`GitlabIssuesParser`	`.json` (auto-detected)	—	Issues, labels, notes, milestones, merge requests

Format auto-detection

The registry detects the format automatically following this order:

Stdin (-): always treated as plaintext
File extension: direct mapping (.md -> markdown, .pdf -> pdf, etc.)
JSON subtype: if the extension is .json, the content is inspected in this order:
- If it has key "issues" or is a list with objects that have "key" + "fields" -> jira
- If it has field "iid" (object or list of objects) -> gitlab_issues
- If it is a list with objects that have "number" + ("html_url" or "labels") -> github_issues
- If it is a list with objects that have "type": "message" + "ts" -> slack
- If no subtype matches -> yaml (treated as structured data)
HTML subtype: if the extension is .html or .htm:
- If the first 2000 characters contain “confluence” or “atlassian” -> confluence
- Otherwise -> fallback to plaintext
URLs: if the source starts with http:// or https:// -> url
Fallback: if there is no parser for the detected format -> plaintext

Note: JSON subtype detection follows a strict priority order: Jira > GitLab Issues > GitHub Issues > Slack > generic YAML. This avoids ambiguities when a JSON has fields that could match multiple formats.

Parsers in detail

Markdown

Extensions: .md, .markdown

What it extracts:

YAML front matter: if the file starts with ---, extracts metadata as key-value pairs
Sections by headings: each #, ##, ###, etc. becomes a section with title, level, and content
Full text: the content without the front matter

Source example:

---
project: User API
version: 2.0
priority: high
---

# Functional Requirements

## FR-01: User registration
The system must allow registration with email and password...

## FR-02: Authentication
The system must support OAuth2 and JWT...

Extracted metadata: project, version, priority (from front matter)

Plain text

Extensions: .txt, stdin (-), files without extension

What it extracts:

Sections by paragraphs: each block separated by blank lines becomes a section
Metadata: source_type (“stdin” or “file”)

Ideal for:

Quick notes
Slack dumps
Raw ideas
Text copied from any source

Example:

We need a real-time notification system.
It must support WebSocket for immediate updates.

Users must be able to configure their preferences:
- Email for important notifications
- Push for real-time updates
- Mute by schedule

YAML / JSON

Extensions: .yaml, .yml, .json (when not Jira)

What it extracts:

Sections by top-level keys: each first-level key becomes a section
Text: YAML representation of the complete content
Metadata: top_level_keys (count) or item_count

Source example:

functional_requirements:
  - id: FR-01
    title: User Registration
    description: Users must be able to register...
    priority: high
    acceptance_criteria:
      - Email validation
      - Password strength check

non_functional_requirements:
  - id: NFR-01
    title: API Response Time
    description: All API endpoints must respond in under 200ms

PDF

Extensions: .pdf Requires: pdfplumber

What it extracts:

Text by page: each page becomes a section
Tables: automatically converted to Markdown format
Metadata: page_count

Limitations:

Only works with PDFs that have extractable text
Scanned PDFs (images only) are not directly supported — use the image parser instead

DOCX

Extensions: .docx Requires: python-docx

What it extracts:

Paragraphs: text from each paragraph
Sections by headings: Word headings are converted into structured sections
Tables: converted to Markdown format
Document metadata: author, title, subject, creation date

Jira

Extensions: .json (auto-detected by structure)

Supports two Jira export formats:

REST API format ({"issues": [...]}):

{
  "issues": [
    {
      "key": "PROJ-001",
      "fields": {
        "summary": "Implement login",
        "description": "The user must be able to...",
        "priority": {"name": "High"},
        "status": {"name": "To Do"},
        "labels": ["auth", "mvp"],
        "comment": {
          "comments": [...]
        },
        "issuelinks": [...]
      }
    }
  ]
}

List format ([{"key": "...", "fields": {...}}, ...]):

[
  {
    "key": "PROJ-001",
    "fields": {
      "summary": "Implement login",
      "description": "..."
    }
  }
]

What it extracts per issue:

Data	Jira field	Limit
Summary	`fields.summary`	—
Description	`fields.description`	—
Priority	`fields.priority.name`	—
Status	`fields.status.name`	—
Labels	`fields.labels`	—
Comments	`fields.comment.comments`	Last 5, max 500 chars each
Issue links	`fields.issuelinks`	Type, direction, target

ADF support: Comments in Atlassian Document Format (nested JSON) are automatically converted to plain text.

Extracted relationships:

blocks / is blocked by
depends on
relates to

Confluence

Extensions: .html, .htm (auto-detected by content) Requires: beautifulsoup4, markdownify

Detection: The first 2000 characters of the file are inspected looking for “confluence” or “atlassian”.

What it extracts:

Main content: searches for the main content div (by id, class, or role)
Markdown conversion: converts HTML to clean Markdown with ATX headings
Sections by headings: from the resulting Markdown
Metadata: title, author, date, description (from <meta> tags)

Content selectors (in priority order):

div#main-content
div.wiki-content
div.confluence-information-macro
div#content
div[role=main]
<body> (fallback)

Images

Extensions: .png, .jpg, .jpeg, .webp, .gif Requires: LLM with vision capability

What it does:

Encodes the image in base64
Sends to the vision LLM with a prompt asking to describe:
- UI mockups / wireframes
- Architecture diagrams
- Visible text in the image
Returns the description as text

Metadata: image_format, file_size_bytes

Note: By default it uses a stub that returns placeholder text. Real vision is activated when the LLMAdapter is configured with a model that supports vision.

URLs

Source: URLs starting with http:// or https:// Requires: httpx, beautifulsoup4, markdownify

What it does:

Downloads the page via httpx (sync, configurable timeout)
Converts HTML to clean Markdown via BeautifulSoup4 + markdownify
Extracts page title, sections by headings
Auto-detects source type by URL patterns

Type auto-detection:

URL pattern	Detected type
`confluence`, `wiki`	confluence
`jira`, `atlassian`	jira
`github.com`	github
Other	webpage

Extracted metadata: url, title, source_type, section_count

Error handling:

Timeout -> ParseError with suggestion to verify the URL
HTTP 4xx/5xx -> ParseError with the status code
Connection error -> ParseError with suggestion to verify the network

Example:

intake init "API review" -s https://wiki.company.com/rfc/auth

Slack

Extensions: .json (auto-detected by structure)

Detection: The JSON file must be a list of objects with "type": "message" and a "ts" field (Slack timestamp).

What it extracts:

Messages: text from each message with user and timestamp
Threads: messages grouped by thread_ts
Decisions: messages with specific reactions (thumbsup, white_check_mark) or keywords like “decided”, “agreed”
Action items: messages with keywords like “TODO”, “action item”, “we need”

Metadata:

Field	Description
`message_count`	Total messages
`thread_count`	Number of threads
`decision_count`	Detected decisions
`action_item_count`	Detected action items

Source example:

[
  {"type": "message", "user": "U123", "text": "We need to use PostgreSQL", "ts": "1700000000.000"},
  {"type": "message", "user": "U456", "text": "Agreed", "ts": "1700000001.000",
   "reactions": [{"name": "thumbsup", "count": 3}]},
  {"type": "message", "user": "U789", "text": "TODO: configure the database", "ts": "1700000002.000",
   "thread_ts": "1700000000.000"}
]

GitHub Issues

Extensions: .json (auto-detected by structure)

Detection: The JSON file must contain objects with a "number" field and at least "html_url", "title" + "labels", or "title" + "body". Supports both a single issue and a list.

What it extracts:

Issues: number, title, body, state (open/closed)
Labels: issue labels
Assignees: assigned users
Milestones: associated milestone
Comments: issue comments
Cross-references: detects #NNN in the text as references to other issues

Supported formats:

// List format (multiple issues)
[
  {
    "number": 1,
    "title": "Login bug",
    "body": "Login fails when...",
    "html_url": "https://github.com/org/repo/issues/1",
    "state": "open",
    "labels": [{"name": "bug"}, {"name": "priority:high"}],
    "comments": [
      {"body": "Reproduced in production", "user": {"login": "dev1"}}
    ]
  }
]

// Individual format (single issue)
{
  "number": 42,
  "title": "Feature request",
  "body": "We need...",
  "html_url": "https://github.com/org/repo/issues/42"
}

Metadata: source_type (“github_issues”), issue_count, labels (comma-separated list), milestone (if exists)

Extracted relationships: cross-references via #NNN in body and comments.

GitLab Issues

Extensions: .json (auto-detected by structure)

Detection: The JSON file must contain objects with an "iid" field (GitLab internal ID). Supports a single issue, a list, or a wrapped format {"issues": [...]}.

What it extracts:

Issues: IID, title, description, state (opened/closed)
Labels: issue labels
Assignees: assigned users
Milestones: associated milestone (title)
Weight: issue weight (if exists)
Task completion status: checkbox progress (count/completed_count)
Discussion notes: non-system discussion notes (max 500 chars each)
Merge requests: linked MRs (as relationships)

Supported formats:

// Individual format (single issue)
{
  "iid": 42,
  "title": "Implement SSO login",
  "description": "Login must support SAML...",
  "state": "opened",
  "labels": ["feature", "auth"],
  "milestone": {"title": "v2.0"},
  "assignees": [{"username": "jdoe"}],
  "notes": [
    {"author": {"username": "dev"}, "body": "Implemented", "system": false}
  ]
}

// List format (multiple issues)
[
  {"iid": 42, "title": "...", ...},
  {"iid": 43, "title": "...", ...}
]

// Wrapped format
{"issues": [{"iid": 42, ...}]}

Metadata: source_type (“gitlab_issues”), issue_count, labels (comma-separated list), milestone (if exists)

Extracted relationships: Linked merge requests (if they exist).

Direct API connectors

In addition to local files, intake can obtain data directly from APIs using scheme URIs. Connectors require configuration in .intake.yaml and credentials via environment variables. There are currently 4 connectors: Jira, Confluence, GitHub, and GitLab.

Jira

Supported URIs:

Pattern	What it does
`jira://PROJ-123`	A single issue
`jira://PROJ-123,PROJ-124,PROJ-125`	Multiple issues
`jira://PROJ?jql=sprint%20%3D%2042`	JQL search
`jira://PROJ/sprint/42`	All issues from a sprint

Dependency: atlassian-python-api (install with pip install "intake-ai-cli[connectors]")

Example:

intake init "Sprint 42 tasks" -s jira://PROJ/sprint/42

Issues are downloaded as temporary JSON and parsed with JiraParser. Comments are included according to connectors.jira.include_comments.

Confluence

Supported URIs:

Pattern	What it does
`confluence://page/123456789`	Page by ID
`confluence://SPACE/Page-Title`	Page by space and title
`confluence://search?cql=space.key%3DENG`	CQL search

Dependency: atlassian-python-api

Example:

intake init "Architecture RFC" -s confluence://ENG/Architecture-RFC-2025

Pages are downloaded as temporary HTML and parsed with ConfluenceParser.

GitHub

Supported URIs:

Pattern	What it does
`github://org/repo/issues/42`	A single issue
`github://org/repo/issues/42,43,44`	Multiple issues
`github://org/repo/issues?labels=bug&state=open`	Issues filtered by labels, state, milestone

Dependency: PyGithub (install with pip install "intake-ai-cli[connectors]")

Example:

intake init "Bug triage" -s github://org/repo/issues?labels=bug&state=open

Issues are downloaded as temporary JSON and parsed with GithubIssuesParser. Maximum 50 issues per query, 10 comments per issue.

GitLab

Supported URIs:

Pattern	What it does
`gitlab://group/project/issues/42`	A single issue
`gitlab://group/project/issues/42,43,44`	Multiple issues
`gitlab://group/project/issues?labels=bug&state=opened`	Filtered issues
`gitlab://group/project/milestones/3/issues`	Issues from a milestone

Dependency: python-gitlab (install with pip install "intake-ai-cli[connectors]")

Example:

intake init "Sprint review" -s gitlab://team/backend/issues?labels=sprint&state=opened

Issues are downloaded as temporary JSON and parsed with GitlabIssuesParser. Maximum 50 issues per query, 10 notes per issue. Supports nested groups and configurable SSL.

Connector configuration

See Configuration > Connectors for the complete configuration fields. Connectors need:

Base URL of the instance (Jira/Confluence) or token (GitHub)
Credentials via environment variables
intake doctor verifies that credentials are configured

General limitations

Limit	Value	Description
Maximum size	50 MB	Files larger than 50 MB are rejected with `FileTooLargeError`
Empty files	Error	Empty files or whitespace-only files produce `EmptySourceError`
Encoding	UTF-8 + fallback	Tries UTF-8 first, fallback to latin-1
Directories	Error	Passing a directory as a source produces an error

Adding support for more formats

There are two ways to add a new parser:

Option 1: Built-in parser (V1 Protocol)

Create a file in src/intake/ingest/ (e.g.: asana.py)
Implement the methods can_parse(source: str) -> bool and parse(source: str) -> ParsedContent
Register it in create_default_registry() and as an entry_point in pyproject.toml

Option 2: External plugin (V2 ParserPlugin)

Create a separate Python package
Implement the ParserPlugin protocol from intake.plugins.protocols
Register as an entry_point in the intake.parsers group in your pyproject.toml

The parser will be automatically discovered when the package is installed. See Plugins for details.

No base class inheritance is required — just implement the correct interface (structural subtyping via typing.Protocol).