Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@ContextCord MCP Server resume my last session and show what changed since the checkpoint"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Engineering continuity and trustworthy context for Coding Agents.
中文 · Website · Quick start · Architecture · Host integration
ContextCord is a local-first continuity and trustworthy-context layer for Coding Agents. It persists engineering state outside individual chats, binds recovered context to repository and source identity, revalidates state when dependencies change, and exposes capability-scoped tools to agent hosts.
5,040 | −84.9% | 299 / 300 |
Engineering records recovered across 60 continuity replays | Bounded Handoff payload vs a 24 KB text baseline | Typed judgments matched the frozen rubric |
Install from a source checkout with Python 3.11+:
python -m pip install -e .Run in your Git project, with at least one commit and an AGENTS.md policy file:
contextcord init --profile generic
contextcord doctor
contextcord feature profile continuity
contextcord handoff --assistHandoff starts with a read-only preview. After closing the preceding session,
confirm with contextcord handoff --assist --task-id <id> --yes to create one
local session and archive its context packet. See the task walkthrough.
Related MCP server: AgentBridge MCP
Module | Responsibility |
Continuum | Task, Session, Checkpoint, Handoff and Closeout lifecycle |
Context Engine | BM25 candidate reduction, freshness, authority and bounded packets |
Veritas | Source identity, dependency fingerprints, object and receipt integrity |
Decision Plane | Deterministic policy, typed decisions and optional route planning |
MCP / Host Gateway | Capability-scoped tools and host adapter/config contracts |
Jev (optional) | Narrow semantic judgments and selection on ambiguous candidate pools |
One repository-local StateStore connects the modules. The code-derived surface has 12 runtime capabilities, 16 MCP tools and 10 typed decision categories. Generated inventory.
Handoff collects local candidates, ranks them with BM25, checks source freshness and authority, and reduces the eligible pool to at most 24 candidates. The default path selects up to three records locally. When explicitly enabled, Jev can rank an ambiguous pool; unavailable or invalid advice uses a recorded BM25 fallback. The resulting packet links source identity, archive references and its decision receipt. Stale records remain marked for revalidation.
Deterministic code owns permissions, integrity, execution controls, tests, destructive approval and release gates. Host permissions and OS isolation govern real execution. Security model.
Measurement | Recorded result |
Task-spec-derived continuity replays | 60 |
Task fields and lossless note sets recovered | 60/60 each; 5,040 notes |
Handoff payload vs bounded-text baseline | 217,067 B / 1,440,000 B; 84.9% reduction |
Dependency-bound records marked for revalidation after source changes | 180; stale selections 0 |
Tampered objects rejected | 60/60 |
Synthetic typed judgments matched the frozen rubric | 299/300; 99.7% |
Live batched typed requests and latency | 30; P95 453 ms |
Deterministic gate replay | Observed bypasses 0/300 |
The continuity replay measures exact state recovery and packet size across 60 task-spec-derived fixtures. The synthetic typed evaluation measures the typed decision contract. All table values can be recomputed from the included rows:
python evaluation/continuity/recompute.py --check
python evaluation/typed_decisions/recompute.py --check --replayEvaluation method and artifacts.
Ships adapter/config contracts for Codex, Cursor, Qoder, OpenCode and WorkBuddy; host qualification is tracked separately from integration support.
Use contextcord feature profile minimal, continuity, decision or full to
select a capability set. MCP mutations require the explicit --allow-mutations
server option. Jev is optional: install .[jev], enable the capability and
configure credentials locally. Profiles and host setup.
Optional model routing and external Model Intelligence provide advisory route planning. Decision Plane.
ContextCord 0.6.2a1 is an alpha release candidate. See CONTRIBUTING.md for verification commands. Distributed under Apache-2.0, with retained third-party notices and acknowledgements.
This server cannot be deployed
Shared project context for AI agents and teams: docs, tasks, and messages that stay current.
Project memory for coding agents: requirements, decisions, code graph and delivery telemetry.
Shared, versioned context that humans and AI agents can publish, review, annotate, and continue.
Project memory, semantic code search, and grounded agent context.