@paperclipai/adapter-openrouter

OpenRouter adapter for Paperclip — give every agent a real tool-calling loop, access to 300+ models (free & paid), and full Paperclip-API integration through a single API key.

If it can receive a heartbeat, it’s hired. Now it can think with any model — and actually do things.

---

What This Does

Connects Paperclip agents to OpenRouter with a complete agent runtime built directly into the adapter — no CLI subprocess, no wrapper, no missing capabilities.

Agent capabilities

• Multi-turn tool-calling loop — model calls tools, adapter executes them, results feed back, loop continues until the model is done or maxTurns is hit
• Built-in Paperclip API tools — 8 tools wired to Paperclip’s REST API so agents can read issues, post comments, update status, create sub-issues, hire teammates, and request approvals
• Auto issue state management — issues move to in_progress when work starts and done / blocked when it finishes
• Final output posted as a comment — every run leaves a comment on the issue so other agents and humans can see the result
• Smart issue checkout — automatically detects when Paperclip’s heartbeat already pre-locked the issue and skips redundant API calls
• Skill loading — drops SKILL.md files into the agent’s system prompt at runtime
• Reasoning support — DeepSeek R1, QwQ, and other thinking models emit thinking transcript entries separately from the final answer
• Approval gating — hire_agent and other mutating tools route through Paperclip’s approval system by default (override with autoApprove: true)
• Repeat-call loop break — detects when a model gets stuck calling the same tool with the same args and breaks the loop with a clear error

Model access

• 300+ models from OpenAI, Anthropic, Google, Meta, Mistral, DeepSeek, Qwen, and more
• 50+ free models — Llama 4 Maverick, Gemma 3, DeepSeek V3, Qwen3 235B, gpt-oss-120b, etc.
• Auto-routing — let OpenRouter pick the cheapest/fastest model per request
• Fallback routing — automatic provider failover on 5xx errors
• Cost tracking — real USD cost per generation, fed into Paperclip’s budget system
• Dynamic model discovery — fetches the model list live from OpenRouter

---

Verified Working

This adapter has been tested end-to-end against a real running Paperclip instance with a real CEO agent on the PanicButton company. Every tool in the suite has been exercised in production runs:

Tested models:

• ✅ openai/gpt-4o-mini — best-in-class function calling, ~$0.003 per multi-tool run
• ✅ openai/gpt-oss-120b:free — reliable function calling, free tier
• ⚠️ stepfun/step-3.5-flash:free — fakes tool calls as XML text, not recommended (model issue, not adapter)

---

Quick Start

1. Get an OpenRouter API Key

Go to and create a key. Free models work with $0 balance, but OpenRouter caps free-tier accounts at 50 requests per day total. Adding $5 in credits unlocks 1000 free-model requests/day and also lets you use paid models.

2. Drop the adapter into Paperclip

# From your Paperclip repo root:
cd packages/adapters
git clone https://github.com/talhamahmood666/paperclip-adapter-openrouter openrouter

3. Apply registry patches

You need to register the adapter in 3 places:

• server/src/adapters/registry.ts — add openrouterAdapter with supportsLocalAgentJwt: true
• server/src/adapters/builtin-adapter-types.ts — add "openrouter" to the type union
• ui/src/adapters/registry.ts + ui/src/adapters/adapter-display-registry.ts — register the UI side

See REGISTRY_PATCHES.md for the exact diffs. Then:

cd ../../..   # back to paperclip root
pnpm install
pnpm -r build

4. Set credentials

Three environment variables are required:

# Your OpenRouter API key (sk-or-v1-...)
export OPENROUTER_API_KEY="sk-or-v1-your-key-here"

# Required for Paperclip to mint agent JWTs that the adapter uses
# to call Paperclip's own API. Any random 32-byte hex string works.
export PAPERCLIP_AGENT_JWT_SECRET="$(openssl rand -hex 32)"

# Where the adapter should reach Paperclip's API. Default is correct
# for local dev.
export PAPERCLIP_API_URL="http://localhost:3100"

Persist them in ~/.bashrc or .paperclip/.env.

5. Hire an agent

In the Paperclip UI → Org Chart → Hire Agent:

1. Adapter Type: OpenRouter
2. Model: any OpenRouter model id, e.g. openai/gpt-4o-mini (recommended for tools) or openai/gpt-oss-120b:free
3. Test Environment: validates your key and lists available models

Then create an issue and assign it to the new agent. Watch the live run view — you’ll see structured tool_call and tool_result entries as the model works.

---

Architecture

packages/adapters/openrouter/ ├── package.json ├── README.md ├── REGISTRY_PATCHES.md └── src/ ├── index.ts # Root metadata, types, OpenRouter constants ├── server/ │ ├── index.ts # Server barrel — execute, sessionCodec, detectModel, listSkills, syncSkills │ ├── execute.ts # Multi-turn tool loop, issue state mgmt, cost tracking, repeat-call protection │ ├── paperclip-api.ts # HTTP client for Paperclip’s REST API (auth via ctx.authToken JWT) │ ├── tools.ts # 8 OpenAI-format tool definitions + handlers │ ├── transcript.ts # Typed TranscriptEntry emitters (init, assistant, thinking, tool_call, tool_result, result) │ ├── skills.ts # Filesystem-based SKILL.md loader │ └── test.ts # Environment diagnostics + dynamic model fetch ├── ui/ │ ├── index.ts │ ├── parse-stdout.ts # Stdout JSON lines → TranscriptEntry[] │ └── build-config.ts # Form values → adapterConfig JSON └── cli/ ├── index.ts └── format-event.ts # Terminal pretty-print for paperclipai run --watch

How a run works

1. Paperclip’s heartbeat dispatcher wakes the agent and calls execute(ctx)
2. Adapter loads skills from disk and prepends them to the system prompt
3. Adapter renders Paperclip’s wake payload as the user message
4. If ctx.authToken is present, adapter constructs a PaperclipApi client and the 8 tool handlers
5. Issue lock acquisition — adapter checks if Paperclip’s heartbeat already stamped the issue’s executionRunId to this run; if yes, skip checkout, otherwise call POST /api/issues/:id/checkout
6. Issue is moved to in_progress
7. Tool loop runs:
   • Call OpenRouter chat/completions with tools array
   • Model returns either text (done) or tool calls
   • Each tool call is executed against Paperclip’s API
   • Results are fed back as role: "tool" messages
   • Loop until no more tool calls, maxTurns is hit, or repeat-call protection trips
8. Final assistant text is posted as a comment on the issue
9. Issue is moved to done (success), blocked (max_turns / error / repeat loop)
10. Adapter fetches the OpenRouter generation cost and returns the full result with usage + costUsd

Built-in tools

All tools call Paperclip’s REST API authenticated as the agent (via ctx.authToken), so every action is attributed correctly in the audit log.

---

Configuration Reference

---

Example Configs

Production-grade CEO on a paid model

{
  "name": "ceo",
  "adapterType": "openrouter",
  "adapterConfig": {
    "model": "openai/gpt-4o-mini",
    "temperature": 0.3,
    "maxTokens": 4096,
    "maxTurns": 30,
    "systemPrompt": "You are the CEO. Define strategy, delegate, and ship."
  }
}

Free-tier agent on gpt-oss-120b

{
  "name": "researcher",
  "adapterType": "openrouter",
  "adapterConfig": {
    "model": "openai/gpt-oss-120b:free",
    "temperature": 0.7,
    "maxTurns": 20
  }
}

Frontier coding agent

{
  "name": "senior-engineer",
  "adapterType": "openrouter",
  "adapterConfig": {
    "model": "anthropic/claude-sonnet-4-6",
    "temperature": 0.0,
    "maxTokens": 16384,
    "maxTurns": 50,
    "route": "fallback"
  }
}

Reasoning agent with auto-routing

{
  "name": "strategist",
  "adapterType": "openrouter",
  "adapterConfig": {
    "model": "openrouter/auto",
    "reasoning": true,
    "transforms": ["middle-out"]
  }
}

---

Cost Tracking

After each completion the adapter queries OpenRouter’s /api/v1/generation endpoint to get the real USD cost and accurate token counts, then returns them in AdapterExecutionResult. Paperclip deducts this from the agent’s monthly budget. At 80% utilization → soft warning; at 100% → agent auto-pauses.

Free models report $0.00 and don’t eat into your budget.

---

Skills

Drop a directory containing a SKILL.md into the skills root and the adapter loads it into the system prompt at runtime: ~/.openrouter-adapter/skills/ ├── customer-research/ │ └── SKILL.md └── crypto-security/ └── SKILL.md

Override the root with adapterConfig.skillsDir per-agent or PAPERCLIP_SKILLS_DIR env var.

v1 loads every skill in the root. Per-agent skill selection (matching Paperclip’s “desired skills” registry) is planned for v3.

---

How It Differs From Other Adapters

---

Known Limitations

The adapter is verified working end-to-end, but there are some real-world rough edges you should be aware of:

Model-specific quirks

• Free models silently rate-limit. OpenRouter caps free-tier accounts at 50 requests per day total across all free models. Heavy testing will exhaust this quickly. Adding $5 in credits unlocks 1000/day. The error is a clear 429 with Rate limit exceeded: free-models-per-day.
• Some free models fake tool calls. stepfun/step-3.5-flash:free is trained on a different tool format and emits ... as literal text instead of structured tool_calls arrays. The adapter cannot recover from this — it’s a model issue. Use openai/gpt-oss-120b:free or openai/gpt-4o-mini instead.
• Reasoning model token counts vary. DeepSeek R1 and QwQ emit reasoning tokens that don’t always show up in OpenRouter’s usage report. Cost tracking falls back to the /generation endpoint for accuracy.

Adapter-specific limitations (deferred to v3)

• No token streaming inside the tool loop. v1 is non-streaming throughout for reliability — most free models stream tool calls poorly. Streaming is planned for v3.
• No async approval-callback resume. When hire_agent creates an approval, the run completes; it does not pause and wait for the human decision. The approval is processed when the agent next wakes.
• No multimodal / vision attachment handling. Issues with image attachments are passed as text descriptions only. Vision support is planned for v3.
• No per-agent desired-skill filtering. v1 loads every skill in the root directory. Per-agent skill selection (matching Paperclip’s skill registry) is planned for v3.
• No per-model capability detection. The adapter sends tools to every model. Models that don’t support function calling will fail or hallucinate. v3 will probe model capabilities and skip the tools array when unsupported.
• Repeat-call detection is conservative. The adapter breaks the loop after 3 identical consecutive calls. This protects against runaway loops but may also fire on legitimate retries — adjust maxTurns if needed.

Paperclip integration notes

• Paperclip core needs supportsLocalAgentJwt: true on the openrouter adapter entry in server/src/adapters/registry.ts. Without this, ctx.authToken will be undefined and tool calls will be disabled. See REGISTRY_PATCHES.md.
• Three env vars are required, not just the API key. OPENROUTER_API_KEY, PAPERCLIP_AGENT_JWT_SECRET, and PAPERCLIP_API_URL must all be set in the same shell that runs pnpm dev:server.
• Issue checkout has been observed to “fail” cosmetically when expectedStatuses doesn’t include the issue’s current status. The adapter’s default list (backlog, todo, in_progress, in_review, blocked) covers all pre-terminal statuses. If Paperclip ever adds new ones, update paperclip-api.ts.

---

Roadmap

v1 — MVP (one-shot, no tools)

• Basic OpenRouter integration ✅
• Streaming text generation ✅
• Cost tracking ✅

v2 — Current release ✅

• Multi-turn tool loop ✅
• 8 built-in Paperclip API tools ✅
• Issue state management ✅
• Issue checkout with smart pre-lock detection ✅
• Final-output-as-comment ✅
• Skill loading ✅
• Reasoning support ✅
• sessionCodec, detectModel, listSkills, syncSkills ✅
• Cost tracking via Generation API ✅
• Repeat-call loop break ✅
• Approval gating for hire_agent and friends ✅
• End-to-end verified against live Paperclip ✅

v3 — Planned

• Token streaming inside the tool loop
• Async approval-callback resume (pause run → wait → continue)
• Per-agent desired-skill filtering (integrate with Paperclip’s skill registry)
• Multimodal attachment support (vision models)
• Per-model capability detection (skip tools for models that can’t handle them)
• Per-model rate-limit backoff for free tier
• getQuotaWindows integration with OpenRouter /key
• Compaction strategy for long conversations

---

Contributing

PRs welcome. To work on the adapter:

git clone https://github.com/talhamahmood666/paperclip-adapter-openrouter
cd paperclip-adapter-openrouter
pnpm install
pnpm typecheck
pnpm build

Run the test suite (when one exists):

pnpm test

To test against a live Paperclip instance, drop the cloned repo into packages/adapters/openrouter of your Paperclip checkout, apply the registry patches, and trigger a run.

---

License

MIT — same as Paperclip.