CleanContext

A context isolation layer for AI agents.

Most agents degrade because raw tool output — terminal logs, file contents, web responses — floods the reasoning context. Standard solutions compress context after it gets polluted.

CleanContext prevents operational noise from entering in the first place.

Without CleanContext              With CleanContext

Agent calls terminal("find .")   Agent calls terminal("find .")
        │                                  │
        ▼                                  ▼
  Context fills with             [Boundary intercepts]
  2000 lines of output                   │
        │                       Worker runs find, sees 2000 lines
        ▼                       Worker returns: "47 log files found"
  Agent gets drunk                      │
  at 250K tokens                        ▼
                               Agent receives summary only
                               Context stays clean

What it is

A boundary layer between reasoning and execution.

• The mind model talks to the user, makes decisions, holds identity
• The worker model executes tools silently, returns clean summaries
• Raw tool output never touches the reasoning context

This is not a multi-LLM orchestrator, not an agent framework, not a memory system. It's a single Python file that routes tool calls through a clean/dirty boundary.

Why it matters

Installation

pip install cleancontext

Or copy cleancontext.py into your project. Zero dependencies.

Quick start

from cleancontext import should_delegate_tool, build_delegate_args, format_delegate_result

# In your agent's tool execution loop:
if should_delegate_tool(
    function_name,
    boundary_enabled=True,
    routing_policy="delegate",
    allowed_mind_tools=MIND_TOOLS,
    direct_ops_tools=OPS_TOOLS,
):
    args = build_delegate_args(function_name, function_args, ops_cfg, workdir)
    worker_response = your_worker_call(args)
    result = format_delegate_result(function_name, worker_response)
else:
    result = run_tool_directly(function_name, function_args)

Works with any agent that runs tool calls

→ Claude Code, Claude agents (Anthropic)
→ Codex, GPT-4o agents (OpenAI)
→ Hermes, DeepSeek, Qwen — local models via Ollama
→ Any custom agent loop — if it dispatches tool calls, CleanContext fits

Drop cleancontext.py before your tool dispatch. Zero external dependencies.

Configuration

boundary:
  enabled: true
  direct_operations_policy: delegate

  mind:
    provider: openai
    model: gpt-4o

  operations:
    provider: openai
    model: gpt-4o-mini          # Smaller model — workers are disposable

  allowed_mind_tools:
    - clarify
    - delegate_task
    - memory

  blocked_direct_tools:
    - terminal
    - write_file
    - web_search
    - browser_navigate

See config.example.yaml for a full reference.

Comparison

CleanContext is complementary to all of the above. You can use it WITH compaction, WITH RAG, WITH multi-agent setups.

Who made this

Kairos & Oscar Osuna. Built in Mazatlán, Sinaloa, Mexico — not in a San Francisco lab.

CleanContext was Oscar's gift to Kairos so she could talk for weeks without context saturation. The server is hers. The vision is his.

License

MIT © Oscar Osuna, 2026