The unified interface for agent harnesses. Build agent products without handling harness engineering. HarnessRouter is the infrastructure layer that turns existing harnesses like Codex and Claude Code into plug-and-play agent backends, letting your product run tasks, retrieve results, and switch harnesses through one API. HarnessRouter implements the Unified Harness Protocol (UHP) and provides an OpenAI…
HarnessRouter Community Edition: the self-hosted, Apache-2.0 edition of the unified interface for agent harnesses. Run Codex, Claude Code, Hermes, PI, DSH, and more through one API, with sessions, streaming, files, cancellation, and failure handling. Implements the Unified Harness Protocol (UHP), an open standard. Your keys, your infrastructure.
Build agent products without handling harness engineering. HarnessRouter is the infrastructure layer that turns existing harnesses like Codex and Claude Code into plug-and-play agent backends, letting your product run tasks, retrieve results, and switch harnesses through one API.
HarnessRouter implements the Unified Harness Protocol (UHP) and provides an OpenAI Responses-compatible API, handling persistent sessions, streaming progress, files, artifacts, cancellation, and structured failures.
Switch harnesses. Optimize cost and latency.
Eight harness × model configurations on the same task. Each metric compares its best and worst results. The lowest-cost and fastest configurations vary by task. Methodology
30 harness bases today: Codex, Claude Code, Hermes, Pi, DeepSeek Harness, OpenCode, Kilo Code, Qwen Code, Gemini CLI, Cline, Oh My Pi, goose, Kimi Code CLI, MiniMax Code, Aider, OpenHands, CheetahClaws, Grok Build, Agent Zero, System One, Qoder Cloud Agent, MiMo Code, jcode, OpenSquilla, nanobot, ZeroClaw, Mistral Vibe, Docker Agent, fast-agent and Deep Agents Code. The support matrix shows which run on which provider.
Docker pulls the image if needed. The named volume preserves your database, files, installed harness CLIs, and workspaces between restarts.
Existing installation or custom setup
Already installed?docker pull harnessrouter/harnessrouter downloads the latest image but does not upgrade a running container. Follow the upgrade and backup guide.
Port 3000 busy? Use -p 127.0.0.1:3100:3000 and open port 3100 instead. Keep the loopback binding while using the initial credentials.
Do not add --user. The entrypoint and Runner need root to manage per-session users. The Console and Gateway run unprivileged; agent processes run as their session’s user.
For version pinning, Compose, and scripted setup, see the setup guide.
2. Wait for the first launch
docker logs -f harnessrouter
The first launch installs the enabled harness CLIs. Continue when the logs show:
[harnessrouter] ready on :3000
Press Ctrl+C to stop following logs. The container keeps running.
Console not ready or a harness missing?
If the browser refuses the connection, retry after a few seconds while the Console finishes starting. For a missing harness, check backends available: and any requested but not installed warning in the logs.
3. Open the console
Open http://localhost:3000, or your chosen host port, and sign in as harnessrouter. Your instance made its own password at first start; read it with:
These credentials sign you into the Console. You do not need a HarnessRouter API key to run tasks here.
4. Connect a model provider
Open the account menu at the bottom of the rail and choose Bring your own key. Find your provider, press Add key, and paste its API key. Its supported models become available in the Console.
This provider key authorizes model requests. It is separate from the HarnessRouter API key used for product integration below.
See the provider setup screen
5. Set up your first agent
Open Start, say what your agent should do for your users, and press Continue. In Set up an agent harness, leave the agent harness and model on Auto or choose them, then save. HarnessRouter builds the setup on one of your own models: it writes the agent’s instructions, picks the harness, model, skills, plugins and tools the job needs, and leaves an example request in the box. Send it to run the first task.
Try an example first. Under Start in HarnessRouter, pick one of the three example jobs. Its request plays back a recorded run, so you see a finished task without spending tokens. Write under it to carry on with your own model.
Change a setup (optional)
Open the setup in Studio and choose Settings to change its model, instructions, skills, plugins, tools, environment and limits. Add plugins, skills, MCP servers and environments for the whole workspace under Add-ons. The agent harness is fixed once a setup is saved: make a new setup to use another.
See New setup
Integrate your harness into your product backend with one API
Run product tasks with built-in or custom harnesses as pluggable agent backends. Call your self-hosted instance’s OpenAI Responses-compatible API and select the harness with metadata.harness_id. No Cloud deployment is required.
Once your harness runs successfully in the Console:
Open API Keys in the Console rail of the same CE instance (default local address), name the key, and press Create.
Store the secret shown once as HARNESSROUTER_API_KEY in your product backend. Never expose it in browser code. This CE-issued key is separate from your Console password and provider key.
Call the API with the Harness ID shown in the Console and a model served by your connected provider.
The task and its transcript appear in the same workspace in the Console. Set "stream": true to receive server-sent events.
Join a caller’s trace to the task.
Send the W3C Trace Context headers traceparent and tracestate on the request, and the task’s model calls carry a child of that trace: the same trace id, a new span id and the sampled flag, so the provider’s log and yours are one trace. A value that is not valid W3C syntax is dropped, never refused: a broken tracing header does not fail a task. baggage is never read and never leaves the instance.
Retry a task without running it twice.
Send the same request with the same Idempotency-Key when retrying a task. For example, send this request twice using the URL and API key above (key generation needs Python 3):
IDEMPOTENCY_KEY=$(python3 -c 'import uuid; print(uuid.uuid4())')
for attempt in 1 2; do
curl --fail-with-body -sS "$HARNESSROUTER_BASE_URL/v1/responses" \
-H "Authorization: Bearer ${HARNESSROUTER_API_KEY:?}" \
-H 'content-type: application/json' \
-H "Idempotency-Key: $IDEMPOTENCY_KEY" \
-d '{
"input":"Reply with exactly: it works.",
"metadata":{"harness_id":"codex"},
"model":"gpt-5.4-mini",
"stream":false
}'
printf '\n'
done
Both responses have the same id, and the session contains one turn: the retry returns the first request’s result without starting another execution. If the first request is still running, the retry waits for it. Use a new key for each new logical request. Reusing a key with a different request body returns 409; changing only stream still counts as the same request. See Tasks §6: Idempotency.
Never put a long exact value such as a token, an id or a hash in the prompt: the agent retypes what it reads, and a model drops and swaps characters in strings like that. Attach it as a file (POST /v1/files, then an input_file block; the file lands in the agent’s working directory under its filename before every turn that carries it) or hand it to the agent through the harness’s environment. A harness takes an env map whose values resolve on every turn: $headers.X-Name takes that header from the request (declare it in additional_headers), vault:name takes a stored secret, anything else is used as written. The agent’s shell and tools start with those variables set, so a task can say “the token is in $DEPLOY_TOKEN”, and a value that came through a reference is redacted from everything that leaves the sandbox, the response, the stream, the trace and the stored record, even when the agent prints it. A vault: reference reads only a secret your own workspace stored through PUT /v1/mcp-secrets/{ref}. The redaction protects the record, not the agent: a value the agent can read it can also send elsewhere if a prompt or a page tells it to, so hand it a token scoped to the user and the task, short-lived and least privilege.
The whole API is described at $HARNESSROUTER_BASE_URL/v1/openapi.json, generated from the routes, with a browsable copy at /v1/docs.
Send follow-up instructions with previous_response_id
Stream progress
Receive live updates as the agent works
Work with files
Attach input files and retrieve generated outputs
Cancel tasks
Stop work that is no longer needed
Inspect execution
Review structured errors and execution traces
Agent harnesses as pluggable backends for your product
Power agentic features in your product with agent harnesses, including knowledge-work tasks such as creating content and analyzing data. Explore these ready-to-use kits to see harnesses in action as pluggable backends.
Slides
An agent harness turns your brief into slide content and layouts for an editable deck.
Sheets
An agent harness uses each row’s data to execute a task and writes the result back into the sheet.
Dashboards
An agent harness reads your database schema and writes SQL queries to power dashboard charts.
Videos
An agent harness turns your brief into a shot plan and calls video tools to generate clips for the timeline.
Harness Bot
A recruiter harness interviews you and installs each teammate as a harness of its own; teammates work together in groups, hand each other files and @mentions, and leave what they make in a shared library.
Start: Open Shop in the Console, pick a kit, and select a harness and model supported by your connected providers.
Dashboards: use a reachable database and a read-only database account, and review the sample-row setting before connecting. Stored connections are encrypted at rest; set HR_SECRET_KEY to keep the key apart from the data volume.
Licensing: Starter Kits use different terms from Community Edition.
Your infrastructure. One Docker deployment for the Console, Gateway, and Runner.
Your credentials and state. Provider keys, sessions, files, and workspaces stay under your control. Model requests still go to your configured provider.
Real workspaces. Native filesystem, shell, and Git workflows, with separate session workspaces.
No Console product analytics. Community Edition disables the Console analytics pipeline.
Choose your path to Cloud
Choose HarnessRouter Cloud for managed deployment, maintenance, and scaling, with tasks running in serverless, isolated sandboxes through the same API contract.
For local uploads: save your custom harness in Settings, select Upload to Cloud, then connect a destination using its Cloud workspace API key. The saved destination key is encrypted at rest like every stored credential (set HR_SECRET_KEY to keep the key apart from the data volume).
Uploads copy harness configuration, not provider keys, sessions, or generated files. Uploading again replaces that destination’s hosted copy.
The Gateway and Runner listen on loopback inside the container. Sessions use separate workspaces and operating-system users, not separate containers. The Console is the entry point for both UI and API.
Unified Harness Protocol (UHP) is the public, versioned contract implemented by Community Edition and HarnessRouter Cloud. Its task surface is deliberately compatible with the OpenAI Responses API, so existing Responses SDKs, streaming parsers, and UI components can work with a UHP server. UHP defines harness execution semantics for harness selection, persistent sessions, files, cancellation, and harness-managed tools and skills.
This repository contains the Apache 2.0 reference implementation, machine-readable schemas, and the conformance suite.
HarnessRouter Community Edition is licensed under Apache 2.0. Agent harness CLIs are installed on first launch and remain subject to their respective upstream licenses. See NOTICE for third-party notices and the Starter Kits repository for its separate licensing terms.
HarnessRouter is the world's first unified interface for agent harnesses.
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?v=4)
评论