CONTINUUM v1
GitHub LIVE · --:--:--
Features MCP Safety Docs API Contact
open source
cyrax8590@gmail.com
Open source under Apache-2.0 since 2026

AGENTS THAT SURVIVE .

we don't do forgotten context +
verified state

Crashes happen. Context disappears. CONTINUUM remembers what actually matters, so nothing gets done twice.

View Documentation GitHub
675 TESTS PASSING
python 3.11, 3.12, 3.13
9 MCP TOOLS
deny-by-default auth
You

why it matters

SAME CRASH. DIFFERENT RECOVERY.

recovery-compare
100%
V M
most frameworks LOCKED
Full transcript replay, nothing verified.
Replay every tool call after a crash.
Duplicate side effects, silently.
Five calls before a single recovery.
Safe, on brand-ish, instantly forgettable.
VS
pick this one
Continuum RECOVERY
Semantic checkpoints, not conversation dumps.
Resume from a semantic checkpoint, not from scratch.
Idempotent action ledger, zero duplicates.
Less replay. More resuming.
Verified, inspectable, genuinely hard to lose.
Continuum · selected

Same crash, same budget.
Wildly different recovery.

see the features →

agent recovery

NO LOST STATE. NO GUESSES. JUST RECOVERY.

29 event types. 7 recovery modes. 6 checkpoint policies. Every component verified before resume.
FEATURES ● 29 event types

Built for agents that run for hours.

29 event types. 7 recovery modes. 6 checkpoint policies. Hash-chained audit trail. Every component verified before resume.

✓ Verified before resume
CORE CAPABILITIES 6 FEATURES
AGENT RECOVERY SYSTEM
EXPLORE THE CAPABILITIES

Semantic Checkpoints

Not a conversation dump. A compact, versioned tree of verified goals, decisions, findings, evidence, and pending tasks.

Environment Validation

Every checkpoint component checked against the live environment. Staleness propagates through the dependency graph.

Idempotent Action Ledger

7 action states. Two-phase claim/complete protocol. Raises UnknownSideEffect rather than guessing.

Hash-Chained Events

Append-only event log with SHA-256 digests. verify() reports trusted_through for partial tamper recovery.

Recovery Contracts

Deterministic, sealed with integrity hash. Names exactly one next permitted action.

Provenance Tracking

Every state component traces to its origin event. Agent-reported state marked REQUIRES_REVIEW.

ADVANCED RECOVERY
Tamper recovery
Cross-environment validation
Full provenance graph
Automatic checkpoint replay
External-agent state reconciliation
Recovery contract enforcement
START BUILDING →

Built for agents that don't forget.

eleven tools, one protocol

DENY BY DEFAULT. WORKS WITH CLAUDE CODE.

Route external side effects through the ledger without embedding the library.
MCP SERVER

11 tools. Deny-by-default.
Works with Claude Code.

Route external side effects through the ledger without embedding the library.

mcp-server.ts
01record_progress
02checkpoint
03intercept_action
04complete_action
05fail_action

THE SERVER

Route external side effects through the ledger without embedding the library. Mutating tools require caller authorization. Read-only tools stay open to everyone.

TOOLS

continuum_record_progress continuum_checkpoint continuum_intercept_action continuum_complete_action continuum_fail_action continuum_reconcile_action continuum_validate continuum_resume continuum_list_actions

Authorization

Mutating tools match callers against an allowlist from CONTINUUM_MCP_MUTATING_CLIENTS env var or .continuum/mcp-policy.json.

Interactive Fault Recovery Simulator

Simulate real-world agent failures and inspect how CONTINUUM handles recovery contracts.

Select Failure Scenario
continuum-cli v0.1.0

the good stuff

WHAT WE MAKE

One library. Three surfaces: Python SDK, MCP server, CLI.

start building

Up and running in minutes

Not on PyPI yet, install from a clone. The core library and CLI use only the standard library, the mcp extra adds the server.

01

Install

git clone https://github.com/Cyrax321/CONTINUUM
cd CONTINUUM
uv venv
uv pip install -e ".[dev]"

Two entrypoints land on your PATH: continuum (the CLI) and continuum-mcp (the MCP server).

02

Record a run

from continuum import EventType, Run, SQLiteStorage

store = SQLiteStorage("agent.db")
store.create_run(Run(run_id="run_4821", goal="Analyze docs"))
store.append_event("run_4821", EventType.RUN_STARTED, {"total": 10_000})

Append events as the agent works. The chain is hash linked and verifiable after a crash.

03

Resume

continuum resume run_4821
continuum validate run_4821

Decide how a run may safely resume, or validate state against the live environment. See python examples/crash_recovery_agent.py.

Read the full README →
METRICS
675+ TESTS PASSING
14 CLI COMMANDS
9 MCP TOOLS
CAPABILITIES
Semantic Recovery MCP Server State Validation Action Ledger Checkpointing Audit Chain

ask away

THE NOSY SECTION

everything you'd grill us on a call, minus the call.
HOW-IT-WORKS.FRAME
01 What exactly does CONTINUUM do?
ANSWER

It is a recovery layer for long-running AI agents. When an agent crashes, it resumes from a compact semantic checkpoint (a versioned tree of verified goals, decisions, findings, evidence, and pending tasks) rather than replaying the whole conversation, and every component is revalidated against the live environment before it continues.

NEED-FOR-SPEED.FRAME
02 How is this different from just replaying the transcript?
ANSWER

Replay re-runs every tool call and can repeat external side effects. CONTINUUM keeps a semantic state representation, an idempotent action ledger (claim/complete, with UnknownSideEffect raised rather than guessing), and a hash-chained append-only event log. Duplicate work and duplicate charges are refused, not merely hoped away.

THE-FIREHOSE.FRAME
03 What does "verified before resume" mean?
ANSWER

Every checkpoint component is checked against the current environment, and staleness propagates through the dependency graph. Recovery contracts are deterministic and sealed with an integrity hash, naming exactly one next permitted action. Agent-reported progress carries provenance and is marked REQUIRES_REVIEW, so an agent cannot certify its own safety.

THE-WHOLE-THING.FRAME
04 Does it work with my agent framework?
ANSWER

It is framework-agnostic. There are adapters for generic Python, the OpenAI Agents SDK, LangGraph, and LangChain, plus a deny-by-default MCP server (11 tools) that lets Claude Code and other MCP clients route side effects through the ledger without embedding the library.

THE-DO-OVER.FRAME
05 Is it ready to use today?
ANSWER

Phases 1-11 of the roadmap are complete: event log, semantic state, SQLite persistence, checkpointing, validation, the action ledger, the recovery engine, the CLI, crash-recovery examples, and framework adapters. The benchmark suite, cloud API, and dashboard are planned next. Known limitations are tracked openly in STATUS.md and the issue tracker.

open source, contributions welcome

GOOD FIRST ISSUES ARE LABELED

Bug reports, adapters, benchmarks, docs. The issue tracker and CONTRIBUTING.md tell you where to start.

Start contributing