A Go CLI and library for intelligent topic-based context management when chatting with Claude. Instead of naively truncating conversation history when approaching token limits, nostop detects topic shifts, scores relevance, and archives low-relevance topics to SQLite — preserving full message content for later restoration.

• Topic detection — Uses Claude Haiku as a sidecar model to identify conversation topic shifts in real time
• Dynamic context allocation — Current topics receive more context budget; older topics receive proportionally less based on relevance scoring
• Archive, don't compact — Full messages are preserved verbatim in SQLite, not summarized or truncated
• Automatic archival — At 95% context usage, archives lowest-relevance topics until usage drops to 50%
• Topic restoration — Archived topics can be brought back into active context when referenced
• Prompt caching — Integrates with Anthropic's cache control API (5m and 1h TTLs) to reduce costs on repeated system prompts
• Streaming — Full SSE streaming support with callback-based API
• TUI — Four-view Bubbletea interface: Chat, History, Topics, and Debug

• Go 1.24+
• An Anthropic API key

go install github.com/hegner123/nostop/cmd/nostop@latest

Build from source:

git clone https://github.com/hegner123/nostop.git
cd nostop
just build
just install

Set your API key:

export ANTHROPIC_API_KEY="sk-ant-..."

Start the TUI:

nostop

Generate a default config file:

nostop --write-config

nostop searches for config files in order:

1. ./nostop.toml
2. ~/.config/nostop/config.toml
3. ~/.nostop.toml

Key settings in nostop.toml:

[context]
max_tokens = 200000        # Model context limit
archive_threshold = 0.95   # Trigger archival at this usage %
archive_target = 0.50      # Archive until usage drops to this %

[models]
chat = "claude-sonnet-4-20250514"
detection = "claude-3-5-haiku-latest"

nostop can be used as a Go library:

engine, err := nostop.New(nostop.Config{
    APIKey:           os.Getenv("ANTHROPIC_API_KEY"),
    DBPath:           "conversations.db",
    MaxContextTokens: 200000,
    ChatModel:        "claude-sonnet-4-20250514",
    DetectionModel:   "claude-3-5-haiku-latest",
})
defer engine.Close()

conv, _ := engine.NewConversation(ctx, "My Chat", "You are helpful.")
resp, _ := engine.Send(ctx, conv.ID, "Hello!")
fmt.Println(resp.Content)

See CONTRIBUTING.md.

MIT. See LICENSE.