chargehand

OpenSSF Scorecard

chargehand turns a request into a typed Task Spec, runs it on one or more coding-agent sessions (OpenCode or Claude Code), and returns a result contract with evidence for every claim. People call it from a CLI; programs call it over HTTP or MCP.

Status: before 1.0 (the release badge shows the version). Workers are read-only by default; the code preset (since 0.8.0) makes one write a branch that passes its tests in a sandbox. See the changelog for what shipped, the roadmap for what comes next, and benchmarks for how it performs. The guide walks through using it, and what it does, and how we know gives each capability's status with its evidence.

Why

Coding agents answer in prose. A program that calls one needs something it can check: claims tied to files at a commit, diffs, session messages or caller inputs, plus a confidence and the exact prompt chain behind the answer.

chargehand keeps control flow (task graph, budgets, retries, parallelism) in code, not in a prompt. Each worker's context stays small, cached and disposable.

Non-goals: its own agent loop, direct calls to model providers, parallelism for its own sake.

How a run works

1. Intake reads request/v1 and writes a Task Spec with one action.
2. The action decides what happens next:
   • answer runs one worker session.
   • split runs 2–4 read-only subtasks as a task graph; later nodes fork the first node's cached prefix.
   • deny, ask and improve stop and return a reason, questions or an improved request.
   • A preset can require approval above a risk or cost estimate.
3. The evidence resolver checks every claim. Claims that do not resolve move to open_questions.
4. You get result/v1. The run log records tokens, cache hits, cost and the prompt chain of every call.

Presets set tools, budgets and allowed actions: default, cheap, thorough, strict, and draft for program callers that bring their own facts. Optional long-term memory, a cache report per run, Prompt CI that gates prompt changes on paired evals, and a routing report round it out.

Memory and services come from MCP servers you list in the profile's mcp_servers; with none listed, a run is unchanged. A run recalls from any number of memory servers at once, each fact labelled with its source, and with retain on it stores only claims whose citations resolved, with their locators, repository and commit. A preset can give workers read-only tools from a server. chargehand extensions check verifies the setup before a run does. See Memory and services for the Hindsight and Chronicle setups.

Quick start

To run the server from the release images instead, with Docker Compose and no .NET SDK, see Install with Docker Compose (deploy/compose/). To run from a clone, you need the .NET 10 SDK and one worker runtime.

1. Copy profiles/example.json to profiles/local.json and fill in your gateway, models, prices and secret-store item names. Profiles reference secrets by item name and never hold them.
2. Pick a runtime:
   • OpenCode: with opencode on PATH (pinned 2.0.18), no profile block is needed. run, serve and mcp start their own opencode serve on 127.0.0.1 and a free port, with a random password and their own state under chargehand/opencode in the per-user data directory, and stop it on exit. Providers come from the environment variables OpenCode reads (e.g. ANTHROPIC_API_KEY); edit xdg/config/opencode/opencode.json in that directory for more, chargehand never overwrites it. To use a server you run yourself, start it with scripts/opencode-serve.sh <opencode-binary> profiles/local.opencode.json 4296 (config from profiles/opencode.example.json) and add the profile's opencode block (url, password_secret, version); then nothing is started. The script turns off OpenCode's project configuration (a checkout's own opencode.json could start a command); set OPENCODE_DISABLE_PROJECT_CONFIG=1 and OPENCODE_CONFIG_PROJECT_DISABLE=1 if you start the server another way. See ADR 0030 and ADR 0004.
   • Claude Code: with claude on PATH (pinned 2.1.283) and signed in (run claude once and log in), no login setup is needed: workers use the CLI's own login. Placeholder models the profile's models map does not name (all of them with no profile) fall back to the CLI's default model. To use another credential, set one of ANTHROPIC_API_KEY or CLAUDE_CODE_OAUTH_TOKEN (from claude setup-token), not both. The profile's claude_code block (version, binary, and at most one of api_key_secret or oauth_token_secret) overrides that. See ADR 0020.
3. List the directories your repositories live in as repository_roots (/ allows any). A request names a repository and a commit; the worker reads a clone of it at that commit under worker_root, which stays outside the OpenCode user's home directory (ADR 0023, ADR 0003). Without repository_roots, run and mcp also allow the directory they were launched in; serve allows only worker_root (ADR 0028).
4. Run a request:

dotnet run --project src/Chargehand.Cli -- run < request.json

Commands

All commands run as dotnet run --project src/Chargehand.Cli -- <command> and read profiles/local.json.

prompts/ and presets/ come from the current directory when it holds both (this checkout, or /app in the image), otherwise from the ones the build copies next to the binary. The run log is the profile's run_log; unset, it is runs/run-log.jsonl in a checkout and chargehand/run-log.jsonl under the per-user data directory (~/.local/share on Linux, ~/Library/Application Support on macOS) anywhere else.

Prompt CI runs on its own for a pull request that changes prompts/ or presets/: .github/workflows/prompt-ci.yml hands it to a self-hosted runner, which posts the commit status. A fork's pull request or a preset change waits for an approval in the prompt-ci-review environment. The runner has one eval profile per runtime and a default; the label prompt-ci:<runtime> picks another. scripts/prompt-ci.sh <pr-number> runs it by hand (ADR 0019, ADR 0025).

HTTP and MCP

chargehand serve binds 127.0.0.1 and requires Authorization: Bearer <key> on every route. The key comes from the secret-store item named by the profile's http.api_key_secret (ADR 0018). On a private network, http.listen binds another address and http.allowed_hosts names the host clients use; a tag v<Version> publishes the server image ghcr.io/<owner>/chargehand:<Version> with the Claude Code runtime (ADR 0024):

docker run -p <private-ip>:4300:4300 -v <dir-with-profile.json>:/config:ro -e CHARGEHAND_API_KEY=... \
  ghcr.io/<owner>/chargehand:<Version>

MCP clients that opt in to the tasks extension get long runs as tasks. When intake answers with questions, the call comes back as input_required. Outside a task, a call with a progress token gets the run id as a progress notification when the run starts, and a call whose HTTP request carries Prefer: wait=N (at most 60) returns after N seconds as a tool error holding the run id and its run-status/v1, while the run goes on (ADR 0029).

chargehand mcp serves the same tool, tasks and questions over stdio: the client starts the process, and stdout carries only MCP messages, logs go to stderr (ADR 0027). For Claude Code, from a checkout:

claude mcp add chargehand -- dotnet run --project <checkout>/src/Chargehand.Cli -- mcp

From the package

The package is a .NET tool that dnx (.NET 10 SDK) fetches from nuget.org and runs; no install step, no port, no key. Pin <version> to one of its versions. Pass the Claude Code credential as one of CLAUDE_CODE_OAUTH_TOKEN or ANTHROPIC_API_KEY; CHARGEHAND_RUNTIME and CHARGEHAND_PROFILE are optional. If a desktop app does not see your shell's PATH, give the full path to dnx.

Claude Code:

claude mcp add chargehand -e CLAUDE_CODE_OAUTH_TOKEN=<token> -- dnx Chargehand@<version> --yes -- mcp

VS Code, .vscode/mcp.json:

{
  "inputs": [{ "id": "claude-token", "type": "promptString", "description": "claude setup-token", "password": true }],
  "servers": {
    "chargehand": {
      "type": "stdio",
      "command": "dnx",
      "args": ["Chargehand@<version>", "--yes", "--", "mcp"],
      "env": { "CLAUDE_CODE_OAUTH_TOKEN": "${input:claude-token}" }
    }
  }
}

Claude Desktop, claude_desktop_config.json:

{
  "mcpServers": {
    "chargehand": {
      "command": "dnx",
      "args": ["Chargehand@<version>", "--yes", "--", "mcp"],
      "env": { "CLAUDE_CODE_OAUTH_TOKEN": "<token>" }
    }
  }
}

scripts/mcp-smoke.py <dir> runs a locally packed tool (dotnet pack src/Chargehand.Cli -o <dir>) the same way and lists its tools; CI runs it on every change.

Claude Code plugin

/chargehand:change <goal> takes one prompt to a reviewed change on a local branch change/<slug>: chargehand researches the goal with citations checked against the current commit, your session writes the change and runs the tests, chargehand reviews the diff with the review preset, the session fixes what holds (at most 2 fix rounds), and a report lands in .chargehand/reports/<slug>.md as its own commit. Nothing is pushed.

/plugin marketplace add Egoushka/chargehand
/plugin install chargehand@chargehand

The plugin starts chargehand through dnx, so it needs the .NET 10 SDK; the Chargehand package it runs is on nuget.org. To run a checkout of chargehand instead of the package, point a chargehand MCP server at it:

{"mcpServers": {"chargehand": {"command": "dotnet", "args": ["run", "--project", "<checkout>/src/Chargehand.Cli", "--", "mcp"]}}}

With no profile the workers run on the agent CLI's default model. To pick models, set CHARGEHAND_PROFILE to a profile whose models map names them; scripts/change-e2e.sh writes a minimal one. --budget applies to each chargehand call, and a run makes up to four (one research, up to three reviews).

Repository layout

Build and test

dotnet build
dotnet test
dotnet format --verify-no-changes

Contributions: see CONTRIBUTING.md. Security reports: see SECURITY.md.

License

Apache-2.0.