ccontext-mcp — Execution Context for AI Agents

English | 中文 | 日本語

Local-first MCP server that gives agents a shared, durable “execution context” across sessions: Vision (why) · Sketch (static blueprint) · Milestones (timeline) · Tasks (deliverables) · Notes/Refs (knowledge) · Presence (who’s doing what).

🧠 Persistent agent memory • 📋 Agent-native task tracking • 🧹 Built-in hygiene (diagnostics + lifecycle) • ⚡ Batch updates (one call) • 🔒 Local files, zero infra

🖼️ ccontext at a Glance

Files on disk (portable, git-friendly)

your-project/
└── context/
    ├── context.yaml           # vision, sketch, milestones, notes, references (+ embedded contract)
    ├── tasks/
    │   ├── T001.yaml          # deliverable tasks with steps
    │   └── T002.yaml
    ├── presence.yaml          # runtime status (recommend gitignore)
    ├── .ccontext.lock         # lock file (recommend gitignore)
    └── archive/               # auto-archived notes/refs/tasks (optional gitignore)

One call to “load the brain”

get_context() returns version + now + diagnostics so agents can quickly orient:

{
  "version": "abc123def456",
  "now": {
    "active_milestone": { "id": "M2", "name": "Phase 2", "description": "...", "status": "active" },
    "active_tasks": [{ "id": "T001", "name": "Implement auth", "milestone": "M2" }]
  },
  "diagnostics": {
    "debt_score": 2,
    "top_issues": [{ "id": "NO_ACTIVE_MILESTONE", "severity": "info", "message": "No active milestone." }]
  },
  "context": { "...": "vision/sketch/milestones/notes/references/tasks_summary" }
}

Why ccontext? (Pain → Payoff)

The Pain

• Agents forget what they were doing between sessions.

• Multi-agent work becomes N² coordination noise without a shared “source of truth”.

• Context grows unbounded; old notes become misleading; task state drifts.

The Payoff

• Resume instantly: agents always start from the same structured context.

• Coordinate cleanly: presence shows who’s doing what; tasks show what’s actually done.

• Stay sane: diagnostics highlight context debt; ttl-based lifecycle prevents bloat.

✨ What Makes ccontext Different

🗂️ Local-first, Portable
Context is plain YAML in your repo. No DB, no cloud, no lock-in.

📋 Agent-native Structure
Designed around how agents actually work: vision, blueprint, milestones, tasks, notes.

⚡ Low-friction Updates
commit_updates() batches multiple changes in one call (status + task step + note).

🧹 Context Hygiene
get_context() emits diagnostics + top issues so agents know what to fix.

⏳ Lifecycle Built-in
Notes/refs decay by ttl and auto-archive, keeping context fresh.

👥 Presence That Stays Readable
Presence is normalized (single-line, de-duped) by design.

Core Model (The “Contract”)

• Vision: one-sentence north star. Low frequency.

• Sketch: static blueprint only (architecture, strategy, constraints, major decisions).
  Do not put TODO/progress/task lists here.

• Milestones: coarse phases (typically 2–6). Exactly one active at a time.

• Tasks: deliverables with 3–7 steps. If work spans handoffs, it belongs in a task.

• Notes/References: “things we must not forget” + “where to look”.

• Presence: what each agent is doing/thinking right now (keep it short).

This contract is embedded into context.yaml under meta.contract for standalone use.

Installation

Claude Code

# Using uvx (recommended)
claude mcp add ccontext -- uvx ccontext-mcp

# Or using pipx
claude mcp add ccontext -- pipx run ccontext-mcp

Claude Desktop

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "ccontext": {
      "command": "uvx",
      "args": ["ccontext-mcp"],
      "env": { "CCONTEXT_ROOT": "/path/to/your/project" }
    }
  }
}

Other MCP clients / manual

pip install ccontext-mcp
CCONTEXT_ROOT=/path/to/project ccontext-mcp

Root selection: ccontext uses CCONTEXT_ROOT when set; otherwise it uses the current working directory.

Agent Loop (Recommended)

1. Start every run

ctx = get_context()  # call first

2. If missing, set the foundation

update_vision("Ship a reliable X that achieves Y.")
update_sketch("## Architecture\n...\n## Strategy\n...\n## Risks\n...")

3. Keep one milestone active

create_milestone(name="Phase 1: Foundation", description="...", status="active")

4. Track real work as tasks

create_task(
  name="Implement auth",
  goal="Users can sign in and sessions are validated",
  steps=[
    {"name":"Design", "acceptance":"Spec reviewed"},
    {"name":"Implement", "acceptance":"Tests passing"},
    {"name":"Rollout", "acceptance":"Docs updated"}
  ],
  milestone_id="M1",
  assignee="peer-a"
)

5. Update with low friction (one call)

commit_updates(ops=[
  {"op":"presence.set","agent_id":"peer-a","status":"Auth: implementing session validation; checking edge cases"},
  {"op":"task.step","task_id":"T001","step_id":"S2","step_status":"done"},
  {"op":"note.add","content":"Edge case: empty header triggers fallback path","ttl":50}
])

Tools

Version Tracking (ETag-style)

Agents can detect change without guessing:

v = get_context()["version"]
# ... later ...
if get_context()["version"] != v:
    # someone changed context/tasks
    ctx = get_context()

Note: version is semantic. It intentionally ignores notes/refs ttl decay so frequent reads don’t churn the hash.

Diagnostics & Lifecycle (Context Hygiene)

• Diagnostics: get_context() returns diagnostics (including debt_score and top_issues) so agents can keep the context clean.

• TTL-based lifecycle: notes and references decay by 1 each get_context() call and auto-archive when stale, preventing “memory bloat”.

• Presence normalization: agent IDs are canonicalized and de-duped; status is normalized to a single concise line for readability.

Git Recommendations

Most teams prefer:

context/presence.yaml
context/.ccontext.lock
context/archive/

Commit context/context.yaml and context/tasks/ so work survives sessions and can be reviewed.

Works With (and Without) Orchestrators

• Standalone: any MCP-capable agent client can use ccontext directly.

• Orchestrators: tools like CCCC can read/write the same context/ files for multi-agent runtime UX.

• No MCP? You can still read/write the YAML files manually (you just won’t get MCP ergonomics like batch updates and diagnostics).

License

MIT