Polytoken 0.8.17 has been released — a free local-first daemon agent for programming: a persistent background process that manages sessions with different AI providers, executes shell commands, file edits, and searches in the user's environment, and is controlled via CLI and TUI. The new version connected the GPT-6.1 Sol model, learned to continue and repair sessions without being tied to a terminal, and since September 14 the project has released six releases.

What happened

The 0.8.17 release on September 29, 2026 connected the GPT-6.1 Sol model simultaneously for OpenAI and Codex providers, introduced headless session continuation polytoken continue --no-attach, a command to restore broken sessions sessions repair, and smarter model group failover. Two days earlier, version 0.8.16 connected Claude Sonnet 5.5 and support for native language-server for TypeScript 7 projects, and GLM-5.3 is already available for Mistral and DeepInfra providers. Since September 14, when 0.8.9 was released, the project has collected six consecutive releases up to 0.8.17, meaning new models enter the harness catalog almost simultaneously with their appearance at providers. Polytoken is distributed as a Go toolkit: installation is done via go get polytoken or the shell installer get.polytoken.dev, the project is free in the "free as in beer" format and is not tied to a single vendor.

Context

Polytoken positions itself not as "yet another AI agent", but as an engineering layer (harness) around models: key operational tasks of agent development are moved to a persistent local process. Such a daemon-harness is a response to the growing risk of vendor lock-in: in Polytoken all prompts and reminders are open for editing, the set of tools is not canceled by a vendor's decision, and work does not stop when a provider fails, because model groups switch in a specified order taking rate-limit into account. The local daemon itself holds sessions, selects models, and distributes roles within a combination, for example an orchestrator on GLM-5.2 and execution models like GPT 5.6 or DeepSeek-V4-Flash. Judging by the pace of releases since mid-September, local harnesses are being distinguished as a separate layer of the industry, where the model catalog is replenished like a "multiplication table", and competition shifts to the quality of orchestration.

Why this matters for the industry

For construction teams, Polytoken provides ready-made primitives that ML teams usually have to assemble themselves: model groups taking rate-limit into account, headless continuation, session recovery, and a limit on parallel subagents in the daemon. The polytoken continue --no-attach wrapper looks like a minimal base for CI scripts: the agent completes a task in the background, without a live terminal. A vendor-independent layer where prompts are open for editing opens the way to reproducible model comparisons on identical prompts in an identical environment. A broader signal is that the layer of agent harnesses is being commoditized: startups have room in the superstructure over the daemon — model arbitration by price and quality, session observability, CI bots and dashboards, and the question of "which model is inside" is gradually being replaced by the question of "who orchestrates better".

Why this matters for users

You can try it with one command: curl -fsS https://get.polytoken.dev | bash installs the agent on Linux or macOS (amd64/arm64) without sudo, with SHA-256 verification, and updates come via the signed feed polytoken update. In everyday work, this means that when the main model is unavailable, the agent itself switches to the next one in the group, background shell tasks are displayed with live output in the Jobs viewer, the daemon.max_concurrent_subagents limit limits the number of parallel subagents, and the spice_level regulator adjusts the aggressiveness of the model's behavior. A paid assistant can be replaced by a combination of models for a task, for example GLM-5.2 as an orchestrator and GPT 5.6 or DeepSeek-V4-Flash as executors, remaining in a free daemon with your own provider keys.

What is still unknown / limitations

There are no measurements of the quality of agent sessions, benchmarks, or independent reproductions in the sources: the appearance of GPT-6.1 Sol or Claude Sonnet 5.5 in the catalog means only model support, not its verified operation within the harness. The reliability of session repair and failover is confirmed only by release descriptions, not by independent tests. The discussion in the Hacker News thread at the time of preparing the material remained empty, so there are no community reviews yet. Scenarios about the transformation of local daemon-harnesses into a standard infrastructure for agent development on the horizon of one to two years are not confirmed in the source data and require independent verification.

Sources

Author

Look at AI, editorial team