Guide for AI Assistants#

This page is designed for AI assistants and LLMs to understand how to use thinkt tools effectively. For the plain-text version optimized for LLM context windows, see /llms.txt or invoke thinkt docs llms.

What is thinkt?#

thinkt provides unified access to AI coding assistant session data from:

Claude Code (~/.claude/)
Kimi Code (~/.kimi/)
Gemini CLI (~/.gemini/)
Copilot CLI (~/.copilot/)
Codex CLI (~/.codex/)
Qwen Code (~/.qwen/)

Access methods: CLI, REST API, MCP, Go library.

Quick Reference#

Repository Structure#

Path	Description
`/cmd/thinkt/`	CLI application
`/internal/thinkt/`	Core library
`/internal/server/`	HTTP and MCP servers (embeds web + web-lite submodules)
`/docs/hugo/content/`	Documentation
`/internal/server/docs/swagger.yaml`	OpenAPI spec

Key Documentation#

Page	Purpose
CLI Guide	Command-line usage
REST API	HTTP endpoints
MCP Server	MCP tools and setup
Docker	Sandboxed execution

Integration Methods#

1. MCP (Recommended for AI Assistants)#

MCP provides direct tool access for AI assistants.

Tools:

Tool	Purpose	Key Input
`list_sources`	Available sources	none
`list_projects`	All projects	`source?`, `include_deleted?`
`list_sessions`	Sessions for project	`project_id`, `source`
`get_session_metadata`	Session overview	`path`, `summary_only?`
`get_session_entries`	Session content	`path`, `limit`, `offset`, `roles`, `include_thinking`
`list_active_sessions`	Active coding sessions	none
`search_sessions`	Search across indexed sessions (supports regex)	`query`, `project?`, `source?`, `limit?`, `case_sensitive?`, `regex?`
`semantic_search`	Search by meaning using embeddings	`query`, `project?`, `source?`, `limit?`, `max_distance?`, `diversity?`

Best Practices:

Call get_session_metadata with summary_only: true first for quick user-intent previews
Use roles: ["user"] to get just prompts
Use entry_indices to fetch specific entries
Set include_thinking: true only when needed
Default limit is 5; paginate with offset

2. CLI#

thinkt projects                     # List projects
thinkt sessions list -p <path>      # List sessions
thinkt sessions view                # View session content
thinkt web                          # Open web interface (auto-starts server)
thinkt web lite                     # Open lightweight debug interface
thinkt server                        # Start HTTP server (foreground)
thinkt server start                  # Start HTTP server (background)
thinkt server status                 # Check server status
thinkt server stop                   # Stop background server
thinkt search "query"                # Search indexed sessions
thinkt semantic search "query"       # Semantic search
thinkt embeddings list               # List embedding models
thinkt embeddings status             # Embedding status
thinkt apps                         # List open-in apps and terminal capability
thinkt apps set-terminal            # Set default terminal for resume

3. REST API#

Base: http://localhost:8784/api/v1

GET /info                            # Server fingerprint, version, uptime
GET /sources
GET /projects?include_deleted=false
GET /projects/{source}/{id}/sessions
GET /sessions/{path}?limit=10&offset=0
GET /sessions/{path}/metadata?summary_only=true&limit=5
GET /search?q=query                  # Full-text search
GET /semantic-search?q=query         # Semantic search
GET /stats                           # Usage statistics
GET /indexer/status                  # Live indexer status

4. Go Library#

import "github.com/wethinkt/go-thinkt/internal/thinkt"

discovery := thinkt.NewDiscovery(claude.Factory(), kimi.Factory(), gemini.Factory(), copilot.Factory(), codex.Factory(), qwen.Factory())
registry, _ := discovery.Discover(ctx)
projects, _ := registry.ListAllProjects(ctx)

Data Model#

Source (claude|kimi|gemini|copilot|codex|qwen)
  ├── Project (directory path)
  │     └── Session (JSONL file)
  │           └── Entry (message)
  │                 └── ContentBlock (text|thinking|tool_use|tool_result)
  └── Team (multi-agent coordination, Claude Code only)
        ├── Member (agent name + session reference)
        ├── Task (shared task board)
        └── Message (inter-agent inbox)

Entry Roles#

Role	Description
`user`	User messages
`assistant`	AI responses
`tool`	Tool execution
`system`	System messages

Content Block Types#

Type	Fields
`text`	`text`
`thinking`	`thinking`
`tool_use`	`tool_use_id`, `tool_name`, `tool_input`
`tool_result`	`tool_use_id`, `tool_result`, `is_error`

Common Tasks#

Task	Method

| Get user prompts only | MCP: get_session_entries with roles: ["user"] | | Export session | API: GET /sessions/{path} |

Tips#

Metadata first - Check session size before loading full content
Paginate - Sessions can have hundreds of entries
Filter by role - Often only user or assistant messages are needed