Entente

allenchenhan9917starsClaim this server

Entente — provenance engineering for coding agents: handoffs become confirmable, verifiable, retryable contracts; agents form a network with evidence and a way back

Autonomous Agents

Install Run on PlugMCPClaim this server

GitHub Website

Run on PlugMCP

1# Entente

README.mdMCP server

Transport: stdio
Status: Local

Power your agent

Add this server

Install from the publisher. Not a PlugMCP endpoint.

Run on PlugMCP

Host a copy with our keys and logs.

No packaged release in the registry yet. Follow the repository README.

RepositoryPublisher installStep-by-step: add this in Claude

Remote URL or a command on your machine? How the two installs differ

Run on PlugMCP

Host a copy with our keys and logs.

A new server you create. It is not this listing's official endpoint. Use the publisher install when you want their URL and their credentials.

Launch

Server facts

Local

Transport: stdio
License: MIT
Last checked: 2026-10-28

Stars: 17
Tools: —
Uptime: —

Tools

No tool list yet. Remote servers get one after the first health check.

FAQ

Questions

How do I install Entente?

Copy the stdio config for Cursor, Claude Desktop, Claude Code, or VS Code on this page. It uses the publisher's install, not a PlugMCP proxy.

Does Entente need an API key?

No required environment variable is listed on this record.

What transport does Entente use?

Entente runs locally over stdio.

When was Entente last checked?

Last checked 2026-10-28. Entente — provenance engineering for coding agents: handoffs become confirmable, verifiable, retryable contracts; agents form a network with evidence and a way back

From the publisher

Documentation

Clarify before execution — the recipient accepts a Task Contract or asks material questions before starting work.
Check evidence, not just status — relayd runs declared checks in the task worktree and records mismatches with the agent's self-report.
Repair the failed task — a delta repair identifies the failed criteria and requested corrections, with an explicit repair budget.
Keep the delivery history — contracts, checks, human decisions, and repairs become JSONL events that drive state, graph, and replay.

Built at the FUTUREMODE BUILDMODE Gen-AI Hackathon 2026.

Our scope: from prompts to provenance

Entente's scope model: Prompt inside Context, Harness, Loop, Graph, and Provenance Engineering

We propose Provenance Engineering as Entente's outermost scope: preserving where decisions and results came from, which version they concern, what verified them, and whether that evidence still applies. This is our product framing; the layers accumulate rather than replace one another.

Scope, small → large	The question it addresses
Prompt Engineering	How do we express this instruction?
Context Engineering	What information does the agent need?
Harness Engineering	Which tools, permissions, and environment support one agent?
Loop Engineering	How does a task execute, receive feedback, repair, and stop?
Graph Engineering	How do agents, tasks, and dependencies coordinate?
Provenance Engineering — our proposed scope	Where did this delivery come from, and what evidence supports it?

Today: Task Contracts, linting, daemon-executed checks, bounded repair, event replay, and Relay Terminal are implemented. Versioned context checkpoints and a delivery Passport are next-step proposals. A traceable log alone does not prove semantic correctness, and running a check separately does not make its test author independent.

Install

On macOS, Linux, or WSL, install once:

curl -fsSL https://raw.githubusercontent.com/allenchenhan99/entente/main/install.sh | sh

Then open Entente from your own Git project:

cd my-project
entente

The installer downloads the original repository, builds the coordination daemon and native terminal UI, and installs an entente command on your existing PATH. The current directory is your project; installation files live separately under ~/.local/share/entente. Run the same installer again to update. A failed build leaves the previous command working.

Prerequisites: Git, curl, tar, and a C compiler. On macOS, use xcode-select --install; on Ubuntu/WSL, install build-essential and Git first. An existing Node.js 22+ and Rust toolchain are reused. Missing Node.js and Rust are installed privately for Entente. The first native build can take several minutes. If no user-writable command directory is already on PATH, the installer uses sudo only to place the launcher in /usr/local/bin, provided that directory is on PATH. Otherwise, follow the user-owned directory example below.

entente --help
entente status
entente down

ENTENTE_INSTALL_DIR chooses a dedicated, absolute application directory.
ENTENTE_BIN_DIR chooses a writable command directory already on PATH.
ENTENTE_REF selects a branch or tag instead of main.
sh install.sh --no-native builds only the TypeScript daemon and Ink TUI.

For example, use a user-owned directory without sudo:

export PATH="$HOME/.local/bin:$PATH"
curl -fsSL https://raw.githubusercontent.com/allenchenhan99/entente/main/install.sh \
  | ENTENTE_BIN_DIR="$HOME/.local/bin" sh

Keep that directory on your shell's PATH in future terminals. Windows users run the .sh installer inside WSL.

Quick start: no agent or API key

Requirements: Node.js 22+ and Git. The following replay path needs no agent login, daemon, or live model calls after dependencies are installed.

git clone https://github.com/allenchenhan99/entente.git
cd entente
npm ci
npx tsc -b

# Open a recorded run in the terminal UI. Press Ctrl+C to quit.
npx tsx apps/tui/src/index.tsx --replay fixtures/events-live-1.jsonl

Prefer a single command that prints and exits?

npx tsx apps/cli/src/index.ts explain planner --replay fixtures/events-live-4.jsonl

This prints a recorded mission, its six clarification answers, three planned tasks, and integration outcome. Explore the same history through the inbox, a contract, or a task:

npx tsx apps/cli/src/index.ts inbox --replay fixtures/events-live-4.jsonl
npx tsx apps/cli/src/index.ts explain contract:t-auth-routes --replay fixtures/events-live-4.jsonl
npx tsx apps/cli/src/index.ts story --replay fixtures/events-live-4.jsonl --task t-login-page

The completed live-4 fixture has an empty inbox; that is expected. In a live mission, the inbox lists handoffs that need human attention.

Open the two-minute, seven-page report → Team Atrophied Intelligence presents the handoff problem, Contract and context design, terminal Demo, case results, contribution paths, and engineering scope. Diagrams animate automatically; the team Demo video is linked from the deck.

How a handoff works

Clarify before work. Verify before delivery. Repair only what failed.

17-second loop · Open interactive diagram · View static SVG · Diagram source

A Task Contract defines the goal, inputs, constraints, non-goals, allowed paths, acceptance criteria, declared checks, and repair budget. Lint errors block spawning. Checks include command, diff_scope, file_exists, human_review, and llm_judge; their evidence should be interpreted according to how they were produced.

See example contracts, the protocol reference, and the original product design. The live-7 event log records a scope-check failure, a self-report mismatch, a repair request, and eventual mission verification.

Run a live mission

In addition to the quick-start dependencies, configure and log in to a supported runtime (claude and/or codex). Use a disposable demo repository. The commands below use a Bash-compatible shell; the demo initialization script and command checks depend on shell tools. On Windows, use a suitable Bash environment such as WSL for this path.

# Create the demo app as its own Git repository and install its dependencies.
bash demo-repo/scripts/init-demo.sh ../entente-demo
cd ../entente-demo
npm ci
cd ../entente

# Launch using the checked-out CLI; no global npm link is required.
node bin/entente.mjs --repo ../entente-demo

From another terminal in this checkout:

node bin/entente.mjs status --repo ../entente-demo
node bin/entente.mjs down --repo ../entente-demo

The launcher starts or reuses relayd, then opens the TUI. Entente hosts the agent terminals itself.

With a Rust toolchain, build the native terminal daemon and Ratatui client:

cargo build -p termd -p relay-tui
cargo run -p relay-tui -- --replay crates/relay-tui/tests/fixtures/live-7

The launcher selects relayterm when it finds termd and rust when it finds relay-tui; otherwise it uses the TypeScript relay host and Ink TUI. Override with --host, --tui, RELAY_TERMD, or RELAY_TUI as appropriate. The default daemon port is 7420; use --port N to change it, --dir <relayDir> to change local run storage, or --no-spawn to require an existing daemon.

Terminal host	Implementation
`relay`	relayd hosts PTYs with node-pty and serves pane, PTY, and metrics routes in-process
`relayterm`	Rust `termd` hosts PTYs; relayd proxies the same routes

RELAY_CLAUDE_MODEL and RELAY_CODEX_MODEL choose the model passed to spawned runtimes. Without an override, each runtime uses its configured default. Codex agents use an isolated CODEX_HOME, so the user's ordinary Codex configuration is not automatically inherited.

The MCP server is available at /mcp. Task-scoped bootstrap configuration provides the agent's credentials and lifecycle instructions. Recipient tools include relay_get_contract, relay_respond_to_contract, relay_report_progress, relay_submit_evidence, and relay_await_verdict; planner tools include relay_propose_task, relay_revise_task, and relay_ask_human.

What we considered — and did not adopt

Approach	Why we did not select it as the current design
Re-read and summarize all history at every delegation	Repeated work adds cost and handoff latency
One mutable rolling summary shared by all children	Unrelated context accumulates, and concurrent updates can overwrite one another
Entropy as the primary context selector	Information quantity does not establish task relevance; a short requirement can be decisive

These are design tradeoffs, not claims of measured benchmark wins. The current context proposal prioritizes fixed versions, task-relevant handoffs, and end-to-end task validation before cost comparisons. Read #4, the consolidated earlier proposal #6, and research discussion #7. Closing #6 consolidated the discussion; it did not mark the feature implemented.

How to contribute

You can make a useful first contribution without running a live agent. Pick a small, reviewable change:

Contribution	Start here	What a useful submission includes
Improve onboarding	This README and CONTRIBUTING.md	The confusing step, environment, and corrected instructions
Report a reproducible bug	Bug form	Command, expected/actual result, and a sanitized replay when possible
Add a lint rule	Rule proposal, `packages/protocol/src/lint/rules`	One passing contract and one failing contract
Add a demo scenario	Scenario form, `examples`	A clear failure, expected clarification or repair, and replay evidence
Extend a runtime or terminal host	Adapter request, `ports.ts`	An adapter behind the existing port, with injected-fake tests

Fork the repository and create a focused branch.
Describe the problem and expected behavior; discuss changes to protocol or product semantics before implementation.
Run the narrow relevant check, then the integration checks when behavior changes.
Open a PR with what changed, how you verified it, and any remaining limits.

# Example: work on contract linting without launching agents.
npx vitest run packages/protocol/src/lint/lint.test.ts

# TypeScript integration checks (also run by CI).
npx tsc -b
npx vitest run

Read the contribution guide → It maps changes to tests, explains fixtures and generated docs, and includes a first-PR checklist.

Repository map

Area	Responsibility
`packages/protocol`	Zod contracts, events, reducer, graph model, lint rules, public schemas
`apps/relayd`	Orchestration, MCP, HTTP/SSE, checks, repairs, worktrees, runtime and host adapters
`apps/launcher`, `apps/cli`	Launch Entente; inspect and operate missions
`apps/tui`	Ink terminal interface and live/replay clients
`crates/termd`, `crates/relay-tui`	Native PTY host and Ratatui terminal client
`demo-repo`, `examples`, `fixtures`	Demo application, task plans, replayable event logs
`docs`	Protocol reference, implementation plans, and research

The Relay Terminal plan records the Phase 2 web-terminal work package; it is design context, not a claim that every planned UI is shipped.

License

MIT — contributions and reproducible coordination failures are welcome.

See third-party sources, licenses, and data provenance for dependency inventories, runtime terms, assets, and synthetic fixture data.

Entente

allenchenhan9917starsClaim this server

Entente — provenance engineering for coding agents: handoffs become confirmable, verifiable, retryable contracts; agents form a network with evidence and a way back

Autonomous Agents

Install Run on PlugMCPClaim this server

GitHub Website

Run on PlugMCP

1# Entente

README.mdMCP server

Transport: stdio
Status: Local

Power your agent

Add this server

Install from the publisher. Not a PlugMCP endpoint.

Run on PlugMCP

Host a copy with our keys and logs.

No packaged release in the registry yet. Follow the repository README.

RepositoryPublisher installStep-by-step: add this in Claude

Remote URL or a command on your machine? How the two installs differ

Run on PlugMCP

Host a copy with our keys and logs.

A new server you create. It is not this listing's official endpoint. Use the publisher install when you want their URL and their credentials.

Launch

Server facts

Local

Transport: stdio
License: MIT
Last checked: 2026-10-28

Stars: 17
Tools: —
Uptime: —

Tools

No tool list yet. Remote servers get one after the first health check.

FAQ

Questions

How do I install Entente?

Copy the stdio config for Cursor, Claude Desktop, Claude Code, or VS Code on this page. It uses the publisher's install, not a PlugMCP proxy.

Does Entente need an API key?

No required environment variable is listed on this record.

What transport does Entente use?

Entente runs locally over stdio.

When was Entente last checked?

Last checked 2026-10-28. Entente — provenance engineering for coding agents: handoffs become confirmable, verifiable, retryable contracts; agents form a network with evidence and a way back

From the publisher

Documentation

Clarify before execution — the recipient accepts a Task Contract or asks material questions before starting work.
Check evidence, not just status — relayd runs declared checks in the task worktree and records mismatches with the agent's self-report.
Repair the failed task — a delta repair identifies the failed criteria and requested corrections, with an explicit repair budget.
Keep the delivery history — contracts, checks, human decisions, and repairs become JSONL events that drive state, graph, and replay.

Built at the FUTUREMODE BUILDMODE Gen-AI Hackathon 2026.

Our scope: from prompts to provenance

Entente's scope model: Prompt inside Context, Harness, Loop, Graph, and Provenance Engineering

Scope, small → large	The question it addresses
Prompt Engineering	How do we express this instruction?
Context Engineering	What information does the agent need?
Harness Engineering	Which tools, permissions, and environment support one agent?
Loop Engineering	How does a task execute, receive feedback, repair, and stop?
Graph Engineering	How do agents, tasks, and dependencies coordinate?
Provenance Engineering — our proposed scope	Where did this delivery come from, and what evidence supports it?

Install

On macOS, Linux, or WSL, install once:

curl -fsSL https://raw.githubusercontent.com/allenchenhan99/entente/main/install.sh | sh

Then open Entente from your own Git project:

cd my-project
entente

entente --help
entente status
entente down

ENTENTE_INSTALL_DIR chooses a dedicated, absolute application directory.
ENTENTE_BIN_DIR chooses a writable command directory already on PATH.
ENTENTE_REF selects a branch or tag instead of main.
sh install.sh --no-native builds only the TypeScript daemon and Ink TUI.

For example, use a user-owned directory without sudo:

export PATH="$HOME/.local/bin:$PATH"
curl -fsSL https://raw.githubusercontent.com/allenchenhan99/entente/main/install.sh \
  | ENTENTE_BIN_DIR="$HOME/.local/bin" sh

Keep that directory on your shell's PATH in future terminals. Windows users run the .sh installer inside WSL.

Quick start: no agent or API key

Requirements: Node.js 22+ and Git. The following replay path needs no agent login, daemon, or live model calls after dependencies are installed.

git clone https://github.com/allenchenhan99/entente.git
cd entente
npm ci
npx tsc -b

# Open a recorded run in the terminal UI. Press Ctrl+C to quit.
npx tsx apps/tui/src/index.tsx --replay fixtures/events-live-1.jsonl

Prefer a single command that prints and exits?

npx tsx apps/cli/src/index.ts explain planner --replay fixtures/events-live-4.jsonl

This prints a recorded mission, its six clarification answers, three planned tasks, and integration outcome. Explore the same history through the inbox, a contract, or a task:

npx tsx apps/cli/src/index.ts inbox --replay fixtures/events-live-4.jsonl
npx tsx apps/cli/src/index.ts explain contract:t-auth-routes --replay fixtures/events-live-4.jsonl
npx tsx apps/cli/src/index.ts story --replay fixtures/events-live-4.jsonl --task t-login-page

The completed live-4 fixture has an empty inbox; that is expected. In a live mission, the inbox lists handoffs that need human attention.

How a handoff works

Clarify before work. Verify before delivery. Repair only what failed.

17-second loop · Open interactive diagram · View static SVG · Diagram source

Run a live mission

# Create the demo app as its own Git repository and install its dependencies.
bash demo-repo/scripts/init-demo.sh ../entente-demo
cd ../entente-demo
npm ci
cd ../entente

# Launch using the checked-out CLI; no global npm link is required.
node bin/entente.mjs --repo ../entente-demo

From another terminal in this checkout:

node bin/entente.mjs status --repo ../entente-demo
node bin/entente.mjs down --repo ../entente-demo

The launcher starts or reuses relayd, then opens the TUI. Entente hosts the agent terminals itself.

With a Rust toolchain, build the native terminal daemon and Ratatui client:

cargo build -p termd -p relay-tui
cargo run -p relay-tui -- --replay crates/relay-tui/tests/fixtures/live-7

Terminal host	Implementation
`relay`	relayd hosts PTYs with node-pty and serves pane, PTY, and metrics routes in-process
`relayterm`	Rust `termd` hosts PTYs; relayd proxies the same routes

What we considered — and did not adopt

Approach	Why we did not select it as the current design
Re-read and summarize all history at every delegation	Repeated work adds cost and handoff latency
One mutable rolling summary shared by all children	Unrelated context accumulates, and concurrent updates can overwrite one another
Entropy as the primary context selector	Information quantity does not establish task relevance; a short requirement can be decisive

How to contribute

You can make a useful first contribution without running a live agent. Pick a small, reviewable change:

Contribution	Start here	What a useful submission includes
Improve onboarding	This README and CONTRIBUTING.md	The confusing step, environment, and corrected instructions
Report a reproducible bug	Bug form	Command, expected/actual result, and a sanitized replay when possible
Add a lint rule	Rule proposal, `packages/protocol/src/lint/rules`	One passing contract and one failing contract
Add a demo scenario	Scenario form, `examples`	A clear failure, expected clarification or repair, and replay evidence
Extend a runtime or terminal host	Adapter request, `ports.ts`	An adapter behind the existing port, with injected-fake tests

Fork the repository and create a focused branch.
Describe the problem and expected behavior; discuss changes to protocol or product semantics before implementation.
Run the narrow relevant check, then the integration checks when behavior changes.
Open a PR with what changed, how you verified it, and any remaining limits.

# Example: work on contract linting without launching agents.
npx vitest run packages/protocol/src/lint/lint.test.ts

# TypeScript integration checks (also run by CI).
npx tsc -b
npx vitest run

Read the contribution guide → It maps changes to tests, explains fixtures and generated docs, and includes a first-PR checklist.

Repository map

Area	Responsibility
`packages/protocol`	Zod contracts, events, reducer, graph model, lint rules, public schemas
`apps/relayd`	Orchestration, MCP, HTTP/SSE, checks, repairs, worktrees, runtime and host adapters
`apps/launcher`, `apps/cli`	Launch Entente; inspect and operate missions
`apps/tui`	Ink terminal interface and live/replay clients
`crates/termd`, `crates/relay-tui`	Native PTY host and Ratatui terminal client
`demo-repo`, `examples`, `fixtures`	Demo application, task plans, replayable event logs
`docs`	Protocol reference, implementation plans, and research

The Relay Terminal plan records the Phase 2 web-terminal work package; it is design context, not a claim that every planned UI is shipped.

License

MIT — contributions and reproducible coordination failures are welcome.

See third-party sources, licenses, and data provenance for dependency inventories, runtime terms, assets, and synthetic fixture data.