Launch
DeepSeek Harness and Pi are now live on HarnessRouter
Following Codex, Claude Code, and Hermes, HarnessRouter now runs five agent harnesses through one interface. DeepSeek Harness and Pi are supported today.
HarnessRouter is the world's first unified interface for agent harnesses: one managed API that lets a product run, trace, and compare multiple agent harnesses without owning the runtime underneath. Today that interface supports five harnesses instead of three.
Developers increasingly treat agent harnesses as an infrastructure layer for building their own products, not as a single tool to commit to. More supported harnesses means more room to pick the right one per task, and to switch later without rebuilding the backend.
More harnesses. One protocol. One API. The two new harnesses, DeepSeek Harness and Pi, use the same create-and-run contract as the three already supported, so nothing about your integration changes when you add them.
Supported harnesses
Five agent harnesses, one API contract
Each harness keeps its own runtime license and model coverage. HarnessRouter gives all five the same interface.
Two axes matter when you pick a harness: whether its runtime is open source, and whether it can run every model family or only its vendor's own models. Here is where the five stand today.
- Codex (base id codex): open source runtime, Apache-2.0. Through HarnessRouter it runs OpenAI models.
- Claude Code (base id claude-code): proprietary runtime under Anthropic's terms. Through HarnessRouter it runs Claude models.
- Hermes (base id hermes): HarnessRouter's own harness. Runs all model families.
- Pi (base id pi): open source runtime, MIT, from earendil-works. Runs all model families. New today.
- DeepSeek Harness (base id dsh): open source runtime, MIT, developer preview, version-pinned. Runs all model families, with DeepSeek models as the default. New today.
License here means the harness runtime's own license, not the model it calls: the Codex CLI is Apache-2.0 and still calls OpenAI's models. Model families are the ones offered through HarnessRouter today, which you can list per harness at GET /v1/harnesses/{id}/models.
One contract
Call every harness the same way
You create a harness once, then run it. To create, POST to /v1/harnesses with a base, which is where the new ids come in: base dsh for DeepSeek Harness, base pi for Pi. You can also set a name, a default model, a system prompt, MCP servers, and skills.
To run, POST to /{harness_id}/v1/responses with your input, and send an Idempotency-Key on the first request so a retry never starts the work twice. The response streams task, run, session, file, and artifact events your product can render directly.
This is the same contract Codex, Claude Code, and Hermes already use. Switching a task from one harness to another is a change of base and model, not a rewrite. The full create-and-run reference lives in the AGENTS.md document for each harness.
New
DeepSeek Harness: open, in developer preview
DeepSeek Harness is an open source agent harness, MIT licensed, that DeepSeek ships as a developer preview. HarnessRouter runs a version-pinned build so behavior stays stable while the upstream project moves quickly.
It defaults to DeepSeek models but is not limited to them: through HarnessRouter it can run all model families, so you can pair the harness with whichever model gives you the best cost and quality on a given task.
Because it is an upstream developer preview, treat its maturity as DeepSeek describes it, not as a HarnessRouter guarantee. It is ready to run today, and version pinning is how we keep preview changes from surprising your product.
New
Pi: a minimal, steerable coding agent
Pi is an open source coding agent from earendil-works, MIT licensed, designed to be small and steerable with a focused set of core tools. HarnessRouter runs it as a first-class harness on the same interface as the others.
Like Hermes and DeepSeek Harness, Pi can run all model families through HarnessRouter, so its behavior is shaped by the model you route to it and the system prompt and skills you give it.
Why it matters
More harnesses, no lock-in
The point of a unified interface is not any single harness. It is that your product does not have to bet on one. With five supported harnesses, you can start on the one that works today, compare model and harness combinations on your own tasks, and move work to a better setup as it appears.
Every harness runs in a sandbox, streams the same events, and returns the same artifact contracts. HarnessRouter records the traces so the choice of harness and model stays measurable over time rather than a one-time guess.
Models generate tokens. Harnesses complete work. HarnessRouter is the interface and control plane above the work layer, and today that layer is five harnesses wide.
FAQ
DeepSeek Harness and Pi FAQ
What are the base ids for the new harnesses?
DeepSeek Harness uses base dsh and Pi uses base pi. You pass the base when you create a harness with POST /v1/harnesses, then run it at POST /{harness_id}/v1/responses.
Is DeepSeek Harness production ready?
It is an upstream developer preview from DeepSeek, MIT licensed, and HarnessRouter runs a version-pinned build. It is ready to run today, but its maturity is what DeepSeek states for the preview, not a HarnessRouter guarantee. Version pinning keeps upstream changes from reaching your product unexpectedly.
Which models can the new harnesses run?
Both Pi and DeepSeek Harness can run all model families offered through HarnessRouter. DeepSeek Harness defaults to DeepSeek models. You can list the models available for any harness at GET /v1/harnesses/{id}/models.
Do I need to change my integration to use them?
No. DeepSeek Harness and Pi use the same create-and-run contract as Codex, Claude Code, and Hermes. Adding them is a change of base and model, not a rewrite.
Can I self-host these harnesses?
Yes. HarnessRouter Community Edition is Apache-2.0 and ships the same five built-in harnesses, so you can run Codex, Claude Code, Hermes, Pi, and DeepSeek Harness on your own infrastructure.
Build on five agent harnesses through one API.
Run Codex, Claude Code, Hermes, Pi, and DeepSeek Harness with the same tasks, sessions, streaming, files, artifacts, and model by harness traces.
Start building